The problem with smarter robots

Read an article the other week about how Deepmind (at Google) is approaching the training of robotics using simulation, reinforcement learning, elastic weights, knowledge distillation and progressive learning.

It seems relatively easy to train a robot to handle some task like grabbing or walking. But doing so can take an awfully long time. If you want to try to train a robot to grab something and put it someplace. You can have it start out making some random movements of its arm, wrist and fingers (if they have such things) and then use reinforcement learning to help it improve its movements over time.

But if each grab attempt takes 10 seconds, using reinforcement learning may take 10,000 attempts before it starts to make any significant progress and perhaps another 20,000-50,000 more to get expert at it. Let’s see 60K *10 seconds is 10,000 minutes or ~170 hours. And that’s just one object pick and place. But then maybe you would like to grab different parts and maybe place them in different locations. All these combinations start adding up.

And of course doing 1000s of movements will wear out gears, motors, mechanisms etc. If only this could all be done in electronic simulations. Then assuming the simulations are accurate enough the whole thing could be done in a matter of hours without wearing anything out. Enter robot simulators such as NVIDIA Isaac Sim, OpenAI RoboSchool/PyBullet

But the problems with simulation are …

Simulations are getting more accurate but at some point their accuracy defeats its purpose because the real world is always noisy, windy and not as deterministic as any simulation. One researcher said you could conceivable have a two armed robot be trained to throw all of a cell phones components up into the air and they will all land in their proper places, proper orientations. But in the real world this could never actually happen, or if it did, it could only happen once.

Hurricane Ike - 2008/09/12 - 21:26 UTC by CoreBurn (cc) (from Flickr)
Hurricane Ike – 2008/09/12 – 21:26 UTC by CoreBurn (cc) (from Flickr)

Weather researchers have been dealing with this problem in spades for a long time. There appears to be a fundamental limit to how far in advance we can predict weather and it’s due to the accuracy with which sensors operate and the complexity of feedback loops between the atmosphere, oceans, landforms, etc. So at a fundamental level, simulations can never be completely accurate. But they can be better.

Today’s weather simulations we see on TV/radio use models that average a number of distinct simulations, where sensor information has been slightly and randomly modified. Something similar could be done for robotic simulation environments, to make them more realistic.

But there are other problems with training robots to do lots of tasks.

Forget me not…

AI deep learning and reinforcement learning algorithms are great when charged with learning a single task, but having it learn multiple tasks is much harder to do. Because each task requires its own deep neural network (DNN) and if you train a DNN on one task and then try to train in on a another task, it forgets all the learnings from the original task. Researchers call this catastrophic forgetting.

One way researchers have dealt with this problem is to effectively freeze certain DNN nodes from having their weights changed during subsequent training rounds and leave others flexible or changeable. One can see this when one trains an image recognition DNN to classify different objects by importing a well trained object classifier and freezing all of it’s layers except the top one or two and then training these layers to classify new objects.

This works well but you have effectively changed the DNN to forget the original object classification training and replaced it with a new one. One solution to this approach is to have multiple passes of training, after each one, certain nodes and connections (of importance to that particular task) are selectively frozen. This works well for a limited number of different tasks but over time all nodes become frozen which means that no more learning can take place. Researchers call this approach to the catastrophic forgetting problem elastic weights.

One way to get around the all nodes frozen issue in elastic weights is to have multiple NNs. One which is trained on a specific task and whose weights are frozen and then a DNN that exists alongside this one with it’s own initialized set of weights. But which uses the original DNN as part of the new DNN inputs. This effectively includes and incorporates all the previously learned knowledge into the new, combined DNN. This is called Progressive Neural Networks.

In this fashion one progressive DNN can be sequentially trained on any number of tasks each of which ends up providing input to all subsequent task training activity. Such a progressive network never forgets and can use previously learned knowledge on new tasks.

The problem with progressive DNNs is a proliferation of DNN column. one for each trained task. However there are a couple of approaches to shrinking an ensemble of DNN like progressive training creates into one that is simpler and just as effective. One way is to perturb weights in DNN nodes and see how model prediction accuracy is impacted on all its tasks. If accuracy isn’t impacted that much, then that node and all its connections could be deleted from the model with minimal impact on model accuracy.

Another approach is to use one DNN to train another. Sort of like a teacher-student. This is called Knowledge Distilation. Where one DNN is a large network (the teacher) and a smaller (student) network that is trained to mimic the teacher DNN to achieve similar accuracy. This is done by training the smaller student network to match the predictions/classifications of the larger one.

Google researchers have shown that knowledge distillation works best when the gap in the sizes of the two networks (teacher and student) aren’t that large. They have solved this problem by introducing an intermediate step (called teachers assistent). They train this TA first then use the TA to train the student.

In the above graphic, when using a teacher of size 110 and a student of size 8 the resulting accuracy suffers but if one uses an intermediate DNN, with a size 20 the resultant accuracy of the student is much closer to the teacher..

~~~~

So with realistic simulation we can train a robot to do any specific task, all using only compute resources. And using progressive DNN training, a robot could conceivably be trained to do any number of tasks. And with appropriate knowledge distillation one can reduce the DNN from progressive training into something much smaller (<10%) than the original DNN.

Want a personal robot that can clean up around your place, do the wash, cook your food and do anything else needed. You know what to do.

Math’s war on Gerrymandering

Read an article the other day in MIT Technology Review, Mathematicians are deploying algorithms to stop gerrymandering, that discussed how a bunch of mathematicians had created tools (Python and R applications, links below) which can be used to test whether state redistricting maps are fair or not.

The best introduction to what these applications can do is in a 2021 I. E. Block Community Lecture: Jonathan Christopher Mattingly captured on YouTube video.

For those outside the states, US Congress House of Representatives are elected via state districts. The term gerrymandering was coined in 1812 over a district map created for Massachusetts that took the form of a dragon which all but assured that the district would go Democrat-Republican vs. Federalist in the next election, (source: Wikipedia entry on Gerrymandering). This re-districting process is done every 10 years in every stateafter the US Census Bureau releases their decennial census results which determines the number of representatives that each state elects to the US House of Representatives.

Funny thing about gerrymandering, both the Democrats and the Republicans have done this in the past and will most likely do so in the future. Once district maps are approved they typically stay that way until the next census.

Essentially, what the mathematicians have done is create a way togenerate a vast number of district maps, under a specific series of constraints/guidelines and then can use these series of maps to characterize the democrat-republican split of a trial election based on some recent election result.

One can see in the above show the # of Democrats that would have been elected using the data from the 2012 and 2016 election results under four distinct maps vs a histogram representing the distribution of all the maps the mathematician’s systems created. The four specific maps indicated on the histograms are.

  • NC2012- thrown out by the state courts as being unfair
  • NC2016– one that NC legislature came out with also thrown out as being unfair as it’s equivalent to the NC2012 map
  • Bipartisan Judges – one that a group of independent judges came out with, and
  • Remedial=NC2020 – one submitted by the mathematicians which they deemed “fairer”
These North Carolina congressional district maps illustrate how geometry is not a fail-safe indicator of gerrymandering. The NC 2012 map, with its bizarre district boundaries, was deemed by the courts to be a racial gerrymander. The replacement, the NC 2016 map, looks quite different and tame by comparison, but was deemed to be an unconstitutional political gerrymander. Analysis by Duke’s Jonathan Mattingly and his team showed that the 2012 and 2016 maps were politically equivalent in their partisan outcomes. A court-appointed expert drew the NC 2020 map.

The algorithms (in Python, GitHub repo for GerryChain and R, GitHub repo for redist) take as input an election map which is a combined US census blocks (physical groupings of population defined by US Census Bureau) and some prior election results (that provide the democrat-republican votes for a particular election within those census blocks). And using this data and a list of districting constraints, such as, compactness requirements, minimal breaks of counties (state political units), population equivalence, etc. and using these inputs, the apps generate a multitude or ensemble of district maps for a state. FYI, districting constraints differ from state to state.

Once they have this ensemble of district maps that adhere to the states specified districting constraints one can compare any sample districting map to the histogram and see if a specific map is fair or not. “Fairness” means that it would result in the same Democrat-Republican split that occurred from the highest number of ensemble maps.

The latest process by which districting maps can be created is documented in a research article, Recombination: A Family of Markov Chains for Redistricting. But most of the prior generations all seem to use a tree structure together with a markov chain approach. At the leaves of the tree are the census blocks and the tree hierarchy algorithmically represents the different districting hierarchies.

Presumably, a Markov Chain encodes a method to represent the state’s districting constraints. And what the algorithm does is traverse the tree of census blocks, using the markov chain and randomness, to create district maps by splitting a branch of the tree (=district) off somewhere in the hierarchy, above the census blocks.

Doing this randomly, over a number of iterations, provides a group or ensemble of proper districting maps that can be used to build the histogram for a specific election result. (Suggest reading the above research report for more information on how this works).

One can see the effect of different election results on the distribution of election results that would have occurred with the current ensemble of maps. For example, USH12 uses the census block voting results from the North Carolina, US (Congressional) House election of 2012 and the GOV16 uses the results from the North Carolina Governor election of 2016

It’s somewhat surprising that the US Supreme Court has ruled that districting is a states issue and not subject to constitutional oversight. Not sure I agree but I’m no constitutional scholar/lawyer. So, all of the legal disputes surrounding state’s re-districting maps have been accomplished in state courts.

But what the mathematicians have done is provide the tools needed to create a multitude of districting maps and when one uses prior election results at the census block level, one can see whether any new re-districting map is representative of what one would see if one drew 100s or 1000s of proper redistricting maps.

Let’s hope this all leads to fairer state and federal elections in the future.

Comments?

Photo Credits:

CTERA, Cloud NAS on steroids

We attended SFD22 last week and one of the presenters was CTERA, (for more information please see SFD22 videos of their session) discussing their enterprise class, cloud NAS solution.

We’ve heard a lot about cloud NAS systems lately (see our/listen to our GreyBeards on Storage podcast with LucidLink from last month). Cloud NAS systems provide a NAS (SMB, NFS, and S3 object storage) front-end system that uses the cloud or onprem object storage to hold customer data which is accessed through the use of (virtual or hardware) caching appliances.

These differ from file synch and share in that Cloud NAS systems

  • Don’t copy lots or all customer data to user devices, the only data that resides locally is metadata and the user’s or site’s working set (of files).
  • Do cache working set data locally to provide faster access
  • Do provide NFS, SMB and S3 access along with user drive, mobile app, API and web based access to customer data.
  • Do provide multiple options to host user data in multiple clouds or on prem
  • Do allow for some levels of collaboration on the same files

Although admittedly, the boundary lines between synch and share and Cloud NAS are starting to blur.

CTERA is a software defined solution. But, they also offer a whole gaggle of hardware options for edge filers, ranging from smart phone sized, 1TB flash cache for home office user to a multi-RU media edge server with 128TB of hybrid disk-SSD solution for 8K video editing.

They have HC100 edge filers, X-Series HCI edge servers, branch in a box, edge and Media edge filers. These later systems have specialized support for MacOS and Adobe suite systems. For their HCI edge systems they support Nutanix, Simplicity, HyperFlex and VxRail systems.

CTERA edge filers/servers can be clustered together to provide higher performance and HA. This way customers can scale-out their filers to supply whatever levels of IO performance they need. And CTERA allows customers to segregate (file workloads/directories) to be serviced by specific edge filer devices to minimize noisy neighbor performance problems.

CTERA supports a number of ways to access cloud NAS data:

  • Through (virtual or real) edge filers which present NFS, SMB or S3 access protocols
  • Through the use of CTERA Drive on MacOS or Windows desktop/laptop devices
  • Through a mobile device app for IOS or Android
  • Through their web portal
  • Through their API

CTERA uses a, HA, dual redundant, Portal service which is a cloud (or on prem) service that provides CTERA metadata database, edge filer/server management and other services, such as web access, cloud drive end points, mobile apps, API, etc.

CTERA uses S3 or Azure compatible object storage for its backend, source of truth repository to hold customer file data. CTERA currently supports 36 on-prem and in cloud object storage services. Customers can have their data in multiple object storage repositories. Customer files are mapped one to one to objects.

CTERA offers global dedupe, virus scanning, policy based scheduled snapshots and end to end encryption of customer data. Encryption keys can be held in the Portals or in a KMIP service that’s connected to the Portals.

CTERA has impressive data security support. As mentioned above end-to-end data encryption but they also support dark sites, zero-trust authentication and are DISA (Defense Information Systems Agency) certified.

Customer data can also be pinned to edge filers, Moreover, specific customer (director/sub-directorydirectories) data can be hosted on specific buckets so that data can:

  • Stay within specified geographies,
  • Support multi-cloud services to eliminate vendor lock-in

CTERA file locking is what I would call hybrid. They offer strict consistency for file locking within sites but eventual consistency for file locking across sites. There are performance tradeoffs for strict consistency, so by using a hybrid approach, they offer most of what the world needs from file locking without incurring the performance overhead of strict consistency across sites. For another way to do support hybrid file locking consistency check out LucidLink’s approach (see the GreyBeards podcast with LucidLink above).

At the end of their session Aron Brand got up and took us into a deep dive on select portions of their system software. One thing I noticed is that the portal is NOT in the data path. Once the edge filers want to access a file, the Portal provides the credential verification and points the filer(s) to the appropriate object and the filers take off from there.

CTERA’s customer list is very impressive. It seems that many (50 of WW F500) large enterprises are customers of theirs. Some of the more prominent include GE, McDonalds, US Navy, and the US Air Force.

Oh and besides supporting potentially 1000s of sites, 100K users in the same name space, and they also have intrinsic support for multi-tenancy and offer cloud data migration services. For example, one can use Portal services to migrate cloud data from one cloud object storage provider to another.

They also mentioned they are working on supplying K8S container access to CTERA’s global file system data.

There’s a lot to like in CTERA. We hadn’t heard of them before but they seem focused on enterprise’s with lots of sites, boatloads of users and massive amounts of data. It seems like our kind of storage system.

Comments?

Photonic [AI] computing seeing the light of day – part 2

Read an interesting article in Analytics India Magazine (MIT Researchers Make New Chips That Work On Light) about a startup out of MIT focused on using photonics for AI/ML/DL activities. Not exactly neuromorphic chips, but using analog photonics interactions to perform computational intensive operations required by todays deep neural net training.

We’ve written about photonics computing before ( see Photonic computing seeing the light of day [-part 1]). That post was about spin outs from Princeton and MIT back in 2019. We showed a bit more on how photonics can perform multiplication and other computations with less power.

The article (noted above) talked about LightIntelligence, an MIT spinout/ startup that’s been around since ~2017, but there’s another company in the same space, also out of MIT called LightMatter that just announced early access to their hardware system.

The CEOs of both companies collaborated on a paper (#1&2 authors of the 10 author paper) written back in 2017 on Deep Learning with Coherent Nanophotonic Circuits. This seemed to be the event that launched both companies.

LightMatter just received $80M in Series B funding ( bringing total funding to $113M) last month and LightIntelligence seems to have $40M in total funding So both have decent funding but, LightMatter seems further ahead in funding and product technology.

LightMatter

LightMatter Envise Photonics-RISC AI processing chip

LightMatter Envise AI chip uses standard RISC electronic cores together with Photo Arithmetic Units for accelerated AI computations. Each Envise chip has 500MB of SRAM for large models, offers 400Gbps chip to chip interconnect fabric, 256 RISC cores, a Graph processor, 294 photonic arithmetic units and PCIe 4.0 connectivity.

LightMatter has just announced early access for their Envise AI photonics server. It’s an 4U, AI server with 16 Envise chips, 2 AMD EPYC CPUs, (16×400=)6.4Tpbs optical fabric for inter-chip communications, 1TB of DDR4 DRAM, 3TB of NVMe SSD and supports 2-200GbE SmartNICs for outside communications.

Envise also offers Idiom Software that interfaces with standard AI frameworks to transform models for photonics computing to use Envise hardware . Developers select Envise hardware to run their AI models on and Idiom automatically re-compiles (IdCompile) their model into more parallelized, photonics operations. Idiom also has a model profiler (IdProfiler) to help debug and visualize photonic models in operation (training or inferencing?) on Envise hardware. Idiom also offers an AI model library (IdML) which provides a PyTorch frontend to help compress and quantize a standard set of AI models.

LightMatter also announced their Passage optical interconnect chip that supplies 100Tbps optical switch for photonics, CPU or GPU processing. It’s huge, 8″x8″ and built on 5nm/7nm node process. Passage can connect up to 48 photonics, CPU or GPU chips that are built onto of it (one can see the space for each of these 48 [sub-]chips on the chip). LightMatter states that 40 Passage (photonic/optical) lanes are the width of one optical fibre. Passage chips are sampling now.

LightMatter Passage photonics-transistor chip (carrier) that provides a photonics programmable interconnect for inter-[photonics-electronic-]chip communications.

LightIntelligence

They don’t appear to be announcing any specific hardware just yet but they are at work in creating the world largest integrated photonics processing system. But LightIntelligence have published a number of research papers focused on photonic approaches to CNNs, RNNs/LSTMs/GRUs, Recurrent ISING machines, statistical computing, and invisibility cloaking.

Turns out the processing power needed to provide invisibility cloaking is very intensive and as its all pixels, photonics offers serious speedups (for invisibility, see Nature article, behind paywall).

Photonics Recurrent ISLING Sampler (PRIS)

LightIntelligence did produce a prototype photonics processor in 2019. And they believe the will have de-risked 80-90% of their photonics technology by year end 2021.

If I had to guess, it would appear as if LightIntelligence is trying to re-imagine deep learning taking a predominately all photonics approach.

Why photonics for AI DL

It turns out that one can use the interaction/interference between two light beams to perform matrix multiplication and other computations a lot faster, with a lot less power than using standard RISC (or CISC) electronic processor architectures. Typical GPUs run 400W each and multi-GPU training activities are commonplace today.

The research documented in the (Deep learning using nanophotonics) paper was based on using an optical FPGA which we have talked about before (See Photonics or Optical FPGAs on the horizon) to prototype the technology back in 2017.

Can photonics change the technology underpinning AI or computing?

If by using photonics, one could speed up AI inferencing by 3-5X and do it with 5-6X less power, you might have a market. These are LightMatter Envise performance numbers on ResNet50 with ImageNet and BERT-Base with SQUAD v1.1 against NVIDIA DGX-A100 (state of the art) AI processing system.

The challenge to changing the technology behind multi-million/billion/trillion dollar industry is that it’s not sufficient to offer a product better than the competition. One has to offer a technology that’s better enough to fund the building of a new (multi-million/billion/trillion dollar) ecosystem surrounding that technology. In order to do that it’s got to be orders of magnitude faster/lower power/better so that commercial customers adopt it en masse.

I like where LightMatter is going with their Passage chip. But their Envise server doesn’t seem fast enough to give them enough traction to build a photonics ecosystem or to fund Envise 2, 3, 4, etc. to change the industry.

The 2017 (Deep learning using nanophotonics) paper predicted that an all optical/photonics implementation of CNN would use 3 orders of magnitude less power for small models and that advantage would only go up for larger models (not counting power for data movement, photo detectors, etc.). Now if that’s truly feasible and maybe it takes a more photonics intensive processor to get there, then photonics technology could truly transform the AI or for that matter the computing industry.

But the other thing that LightIntelligence and LightMatter may be counting on is the slowdown in Moore’s law which may inhibit further advances in electronics processing power. Whether the silicon industry is ready to throw in the towel yet on Moore’s law is TBD.

Comments?

Photo Credit(s):

New era of graphical AI is near #AIFD2 @Intel

I attended AIFD2 ( videos of their sessions available here) a couple of weeks back and for the last session, Intel presented information on what they had been working on for new graphical optimized cores and a partner they have, called Katana Graph, which supports a highly optimized graphical analytics processing tool set using latest generation Xeon compute and Optane PMEM.

What’s so special about graphs

The challenges with graphical processing is that it’s nothing like standard 2D tables/images or 3D oriented data sets. It’s essentially a non-Euclidean data space that has nodes with edges that connect them.

But graphs are everywhere we look today, for instance, “friend” connection graphs, “terrorist” networks, page rank algorithms, drug impacts on biochemical pathways, cut points (single points of failure in networks or electrical grids), and of course optimized routing.

The challenge is that large graphs aren’t easily processed with standard scale up or scale out architectures. Part of this is that graphs are very sparse, one node could point to one other node or to millions. Due to this sparsity, standard data caching fetch logic (such as fetching everything adjacent to a memory request) and standardized vector processing (same instructions applied to data in sequence) don’t work very well at all. Also standard compute branch prediction logic doesn’t work. (Not sure why but apparently branching for graph processing depends more on data at the node or in the edge connecting nodes).

Intel talked about a new compute core they’ve been working on, which was was in response to a DARPA funded activity to speed up graphical processing and activities 1000X over current CPU/GPU hardware capabilities.

Intel presented on their PIUMA core technology was also described in a 2020 research paper (Programmable Integrated and Unified Memory Architecture) and YouTube video (Programmable Unified Memory Architecture).

Intel’s PIUMA Technology

DARPA’s goals became public in 2017 and described their Hierarchical Identity Verify Exploit (HIVE) architecture. HIVE is DOD’s description of a graphical analytics processor and is a multi-institutional initiative to speed up graphical processing. .

Intel PIUMA cores come with a multitude of 64-bit RISC processor pipelines with a global (shared) address space, memory and network interfaces that are optimized for 8 byte data transfers, a (globally addressed) scratchpad memory and an offload engine for common operations like scatter/gather memory access.

Each multi-thread PIUMA core has a set of instruction caches, small data caches and register files to support each thread (pipeline) in execution. And a PIUMA core has a number of multi-thread cores that are connected together.

PIUMA cores are optimized for TTEPS (Tera-Traversed Edges Per Second) and attempt to balance IO, memory and compute for graphical activities. PIUMA multi-thread cores are tied together into (completely connected) clique into a tile, multiple tiles are connected within a single node and multiple nodes are tied together with a 8 byte transfer optimized network into a PIUMA system.

P[I]UMA (labeled PUMA in the video) multi-thread cores apparently eschew extensive data and instruction caching to focus on creating a large number of relatively simple cores, that can process a multitude of threads at the same time. Most of these threads will be waiting on memory, so the more threads executing, the less likely that whole pipeline will need to be idle, and hopefully the more processing speedup can result.

Performance of P[I]UMA architecture vs. a standard Xeon compute architecture on graphical analytics and other graph oriented tasks were simulated with some results presented below.

Simulated speedup for a single node with P[I]UMAtechnology vs. Xeon range anywhere from 3.1x to 279x and depends on the amount of computation required at each node (or edge). (Intel saw no speedups between a single Xeon node and multiple Xeon Nodes, so the speedup results for 16 P[I]UMA nodes was 16X a single P[I]UMA node).

Having a global address space across all PIUMA nodes in a system is pretty impressive. We guess this is intrinsic to their (large) graph processing performance and is dependent on their use of photonics HyperX networking between nodes for low latency, small (8 byte) data access.

Katana Graph software

Another part of Intel’s session at AIFD2 was on their partnership with Katana Graph, a scale out graph analytics software provider. Katana Graph can take advantage of ubiquitous Xeon compute and Optane PMEM to speed up and scale-out graph processing. Katana Graph uses Intel’s oneAPI.

Katana graph is architected to support some of the largest graphs around. They tested it with the WDC12 web data commons 2012 page crawl with 3.5B nodes (pages) and 128B connections (links) between nodes.

Katana runs on AWS, Azure, GCP hyperscaler environment as well as on prem and can scale up to 256 systems.

Katana Graph performance results for Graph Neural Networks (GNNs) is shown below. GNNs are similar to AI/ML/DL CNNs but use graphical data rather than images. One can take a graph and reduce (convolute) and summarize segments to classify them. Moreover, GNNs can be used to understand whether two nodes are connected and whether two (sub)graphs are equivalent/similar.

In addition to GNNs, Katana Graph supports Graph Transformer Networks (GTNs) which can analyze meta paths within a larger, heterogeneous graph. The challenge with large graphs (say friend/terrorist networks) is that there are a large number of distinct sub-graphs within the graph. GTNs can break heterogenous graphs into sub- or meta-graphs, which can then be used to understand these relationships at smaller scales.

At AIFD2, Intel also presented an update on their Analytics Zoo, which is Intel’s MLops framework. But that will need to wait for another time.

~~~~

It was sort of a revelation to me that graphical data was not amenable to normal compute core processing using today’s GPUs or CPUs. DARPA (and Intel) saw this defect as a need for a completely different, brand new compute architecture.

Even so, Intel’s partnership with Katana Graph says that even today compute environment could provide higher performance on graphical data with suitable optimizations.

It would be interesting to see what Katana Graph could do using PIUMA technology and appropriate optimizations.

In any case, we shouldn’t need to wait long, Intel indicated in the video that P[I]UMA Technology chips could be here within the next year or so.

Comments?

Photo Credit(s):

  • From Intel’s AIFD2 presentations
  • From Intel’s PUMA you tube video

Swarm learning for distributed & confidential machine learning

Read an article the other week about researchers in Germany working with a form of distributed machine learning they called swarm learning (see: AI with swarm intelligence: a novel technology for cooperative analysis …) which was reporting on a Nature magazine article (see: Swarm Learning for decentralized and confidential clinical machine learning).

The problem of shared machine learning is particularly accute with medical data. Many countries specifically call out patient medical information as data that can’t be shared between organizations (even within country) unless specifically authorized by a patient.

So these organizations and others are turning to use distributed machine learning as a way to 1) protect data across nodes and 2) provide accurate predictions that uses all the data even though portions of that data aren’t visible. There are two forms of distributed machine learning that I’m aware of federated and now swarm learning.

The main advantages of federated and swarm learning is that the data can be kept in the hospital, medical lab or facility without having to be revealed outside that privileged domain BUT the [machine] learning that’s derived from that data can be shared with other organizations and used in aggregate, to increase the prediction/classification model accuracy across all locations.

How distributed machine learning works

Distributed machine learning starts with a common model that all nodes will download and use to share learnings. At some agreed to time (across the learning network), all the nodes use their latest data to re-train the common model and share new training results (essentially weights used in the neural network layers) with all other members of the learning network.

Shared learnings would be encrypted with TLS plus some form of homomorphic encryption that allowed for calculations over the encrypted data.

In both federated and swarm learning, the sharing mechanism was facilitated by a privileged block chain (apparently Etherium for swarm). All learning nodes would use this blockchain to share learnings and download any updates to the common model after sharing.

Federated vs. Swarm learning

The main difference between federated and swarm learning is that with federated learning there is a central authority that updates the model(s) and with swarm learning that processing is replaced by a smart contract executing within the blockchain. Updating model(s) is done by each node updating the blockchain with shared data and then once all updates are in, it triggers a smart contract to execute some Etherium VM code which aggregates all the learnings and constructs a new model (or at least new weights for the model). Thus no node is responsible for updating the model, it’s all embedded into a smart contract within the Etherium block chain. .

Buthow does the swarm (or smart contract) update the common model’s weights. The Nature article states that they used either a straight average or a weighted average (weighted by “weight” of a node [we assume this is a function of the node’s re-training dataset size]) to update all parameters of the common model(s).

Testing Swarm vs. Centralized vs. Individual (node) model learning

In the Nature paper, the researchers compared a central model, where all data is available to retrain the models, with one utilizing swarm learning. To perform the comparison, they had all nodes contribute 20% of their test data to a central repository, which ran the common swarm updated model against this data to compute an accuracy metric for the swarm. The resulting accuracy of the central vs swarm learning comparison look identical.

They also ran the comparison of each individual node (just using the common model and then retraining it over time without sharing this information to the swarm versus using the swarm learning approach. In this comparison the swarm learning approach alway seemed to have as good as if not better accuracy and much narrower dispersion.

In the Nature paper, the researchers used swarm learning to manage the machine learning model predictions for detecting COVID19, Leukemia, Tuberculosis, and other lung diseases. All of these used public data, which included PBMC (peripheral blood mono-nuclear cells) transcription data, whole blood transcription data, and X-ray images.

Swarm learning also provides the ability to onboard new nodes in the network. Which would supply the common model and it’s current weights to the new node and add it to the shared learning smart contract.

The code for the swarm learning can be downloaded from HPE (requires an HPE passport login [it’s free]). The code for the models and data processing used in the paper are available from github. All this seems relatively straight forward, one could use the HPE Swarm Learning Library to facilitate doing this or code it up oneself.

Photo Credit(s):

Towards a better AGI – part 3(ish)

Read an article this past week in Nature about the need for Cooperative AI (Cooperative AI: machines must learn to find common ground) which supplies the best view I’ve seen as to a direction research needs to go to develop a more beneficial and benign AI-AGI.

Not sure why, but this past month or so, I’ve been on an AGI fueled frenzy (at leastihere). I didn’t realize this was going to be a multi-part journey otherwise, I would have lableled them AGI part-1 & -2 ( please see: Existential event risks [part-0], NVIDIA Triton GMI, a step to far [part-1] and The Myth of AGI [part-2] to learn more).

But first please take our new poll:

The Nature article puts into perspective what we all want from future AI (or AGI). That is,

  • AI-AI cooperation: AI systems that cooperate with one another while at the same time understand that not all activities are zero sum competitions (like chess, go, Atari games) but rather most activities, within the human sphere, are cooperative activities where one agent has a set of goals and a different agent has another set of goals, some of which overlap while others are in conflict. Sport games like soccer lacrosse come to mind. But there are other card and (Risk & Diplomacy) board games that use cooperating parties, with diverse goals to achieve common ends.
  • AI-Human cooperation: AI systems that cooperate with humans to achieve common goals. Here too, most humans have their own sets of goals, some of which may be in conflict with the AI systems goals. However, all humans have a shared set of goals, preservation of life comes to mind. It’s in this arena where the challenges are most acute for AI systems. Divining human and their own system underlying goals and motivations is not simple. And of course giving priority to the “right” goals when they compete or are in conflict will be an increasingly difficult task to accomplish, given todays human diversity.
  • Human-Human cooperation: Here it gets pretty interesting, but the paper seems to say that any future AI system should be designed to enhance human-human interaction, not deter or interfere with it. One can see the challenge of disinformation today and how wonderful it would be to have some AI agent that could filter all this and present a proper picture of our world. But, humans have different goals and trying to figure out what they are and which are common and thereby something to be enhanced will be an ongoing challenge.

The problem with today’s AI research is that its all about improving specific activities (image recognition, language understanding, recommendation engines, etc) but all are point solutions and none (if any) are focused on cooperation.

Tit for tat wins the award

To that end, the authors of the paper call for a new direction one that attempts to imbue AI systems with social intelligence and cooperative intelligence to work well in the broader, human dominated world that lies ahead.

In the Nature article they mentioned a 1984 book by Richard Axelrod, The Evolution of Cooperation. Perhaps, the last great research on cooperation that was ever produced.

In this book it talked about a world full of simulated prisoner dilemma actors that interacted, one with another, at random.

The experimenters programmed some agents to always do the proper thing for their current partner, some to always do the wrong thing to their partner, others to do right once than wrong from that point forward, etc. The experimenters tried every sort of cooperation policy they could think of.

Each agent in an interaction would get some number of points for an interaction. For example, if both did the right thing they would each get 3 points, if one did wrong, the sucker would get 1 and the bad actor would get 4, both did wrong each got 1 point, etc.

The agents that had the best score during a run (of 1000s of random pairings/interactions) would multiply for the the next run and the agents that did worse would disappear over time in the population of agents in simulated worlds.

The optimal strategy that emerged from these experiments was

  1. Do the right thing once with every new partner, and
  2. From that point forward tit for tat (if the other party did right the last time, then you do right thing the next time you interact with them, if they did wrong the last time, then you do wrong the next time you interact with them).

It was mind boggling at the time to realize that such a simple strategy could be so effective/sustainable in simulation and perhaps in the real world. It turns out that in a (simulated) world of bad agents, there would be this group of Tit for Tat agents that would build up, defend itself and expand over time to succeed.

That was the state of the art in cooperation research back then (1984). I’ve not seen anything similar to this since.

I haven’t seen anything like this that discusses how to implement algorithms in support of social intelligence.

~~~~

The authors of the Nature article believe it’s once again time to start researching cooperation techniques and start researching social intelligence so we can instill proper cooperation and social intelligence technology into future AI (AGI) systems .

Perhaps if we can do this, we may create a better AI (or AGI) so that both it and we can live better in our world, galaxy and universe.

Comments?

AI inferencing using light alone

Researchers at UCLA have taken a trained DL neural network and implemented it into a series of passive optical only, 3D printed diffraction gratings to perform fashion MNIST object classification. And did the same with a MNIST handwritten digit and ImageNet DL neural network classifiers.

But first please take our new poll:

Experimental testing of 3D-printed D2NNs.(A and B) After the training phase, the final designs of five different layers (L1, L2, …, L5) of the handwritten digit classifier, fashion product classifier, and the imager D2NNs are shown. To the right of the network layers, an illustration of the corresponding 3D-printed D2NN is shown. (C and D) Schematic (C) and photo (D) of the experimental terahertz setup. An amplifier-multiplier chain was used to generate continuous-wave radiation at 0.4 THz, and a mixer-amplifier-multiplier chain was used for the detection at the output plane of the network. RF, radio frequency; f, frequency.

See the article on SlashGear, 3D printed all-optical diffractive deep learning neural network…. The research article is only available on Optical Society of America’s website/magazine (see Residual D2NN: training diffractive deep neural networks via learnable light shortcuts behind hard paywall). However, I did find a follow on article on ArchivX (see Analysis of Diffractive Optical Neural Networks and Their Integration with Electronic Neural Networks) that discussed how to integrate D2NN approaches with an electronic NN to create a hybrid inference engine. And another earlier Science article (see All-optical machine learning using diffractive deep neural networks) that was available which described earlier versions of D2NN technology for MNIST digit classification, fashion MNIST classification and ImageNet object classification.

How does it work

Apparently the researchers trained a normal (electronic based) deep learning neural network on the MNIST, Fashion MNIST and ImageNet and then converted the resultant trained NNs into a set of multiple diffraction grids. They did some computer simulation of the D2NN and once satisfied it worked and achieved decent accuracy, 3D printed the diffraction plates.

All-optical D2NN-based classifiers. These D2NN designs were based on spatially and temporally coherent illumination and linear optical materials/layers. (a) D2NN setup for the task of classification of handwritten digits (MNIST), where the input information is encoded in the amplitude channel of the input plane. (b) Final design of a 5-layer, phase-only classifier for handwritten digits. (c) Amplitude distribution at the input plane for a test sample (digit ‘0’). (d-e) Intensity patterns at the output plane for the input in (c); (d) is for MSE-based, and (e) is softmax- cross-entropy (SCE)-based designs. (f) D2NN setup for the task of classification of fashion products (Fashion-MNIST), where the input information is encoded in the phase channel of the input plane. (g) Same as (b), except for fashion product dataset. (h) Phase distribution at the input plane for a test sample. (i-j) Same as (d) and (e) for the input in (h),  refers to the illumination source wavelength. Input plane represents the plane of the input object or its data, which can also be generated by another optical imaging system or a lens, projecting an image of the object data onto this plane.

In their D2NN, they start with coherent (laser) light in the THz spectrum, used this to illuminate the input plane (I assume an image of the object/digit/fashion accessory) and passed this through multiple plates of diffraction grids onto THz detector which was used to detect the illuminated spot that indicated the classification.

The article in science has a supplementary materials download that show how the researchers converted NN weights into a diffraction grating. Essentially each pixel on the diffraction grating either transmits, refracts, or reflects a light path. And this represents the connections between layers. It’s unclear whether the 5 or 6 plates used in the D2NN correspond to the NN layers but it’s certainly possible.

And to the life of me I can’t understand what they mean by “Residual D2NN”, other than if it means using a trained (residual) NN and converting this to D2NN.

Some advantages of D2NN

3D printing diffraction gratings means anyone/lab could do this. The 3D printers they used had a spatial accuracy of 600 dpi, with 0.1mm accuracy, almost consumer grade 3D printers. In any case, being able to print these in a matter of hours, while not as easy as changing an all digital NN, seems like an easy way to try out the approach.

For example, for the MNIST digit classifier they used a pixel size of 400um and each diffraction layer they created was equivalent to 200X200 neural weights. Which means that 5 layer D2NN could handle about 0.2M neural weights which were completely connected to one another. This meant they could have (200×200)**2*5=8B connections in the MNIST D2NN. In the image classifier, each diffraction layer had 300×300 neural weights. So D2NN’s seem to scale very well.

Being an all passive optical device, the system is operates entirely in parallel, That is, the researchers indicated that the D2NN devices operate at the speed of light and would perform the inferencing activity in the time it takes a camera to capture the image.

Also the device uses very little energy (I assume just the energy for the THz generator, the input plane detector and the THz detector at the end.

And the researchers also claimed the device was cheap to manufacture, it could be created for less than $50. (Unclear if this included all the electronics or just the D2NN diffraction gratings and holder). And once you have locked into a D2NN that you wanted to use, could be manufactured in volume, very cheaply (sort of like stamping out CD platters). Finally, the number of neural network nodes and layers can be scaled up to a large number of layers and nodes per layer while still fitting on the diffraction gratings. In contrast, all electronic NN require more compute power as you scale up network layers and nodes per layer.

The other article (ArchivX) talked about potentially using a hybrid optical-electronic DNN approach with some layers being D2NN and others being purely digital (electronics). Such a system could potentially be used where some portion of the NN was more stable/more compute intensive than others and where the final output classification layer(s) was more changeable and much smaller/less compute intensive. Such a hybrid system could make use of the best of of the all optical D2NN to efficiently and quickly compress the input space and then have the electronic final classification layer provide the final classification step.

The Oracle

Combining a handful of D2NNs into a device that accepts speech input and provides speech output with the addition of say an offline copy of Wikipedia, Google Books etc. with a search engine that could be used to retrieve responses to questions asked would create an oracle device. Where you would ask a question and the device would respond with the best answer it could find (in it’s databases).

If this could be made out of an all passive optical components and use natural sunlight/electronic illumination to perform it’s functionality, such an all optical, question to answer oracle would be very useful to the populations of the world. And could be manufactured in volume very cheaply and would cost almost nothing to operate.

A couple of other tweaks, if we could collapse the multiple grating D2NNs into a single multi-layer plate/platter and make these replaceable in the device that would allow the oracle’s information base to be updated periodically.

Then if we could embed such a device into a Long Now Clock that would reflect sunlight onto the disk every Solstice, or Equinox, then we could have a quarterly oracle device that could last for 1000 of years. That would provide answers to queries one day every quarter. And that would be quite the oracle…

Photo credit(s):