Despite significant progress in machine learning over the past few years, today’s state of the art algorithms are brittle and do not generalize well. In contrast, the brain is able to robustly separate and categorize signals in the presence of significant noise and non-linear transformations, and can extrapolate from single examples to entire classes of stimuli. This performance gap between software and wetware persists despite some correspondence between the architecture of the leading machine learning algorithms and their biological counterparts in the brain, presumably because the two still differ significantly in the details of operation. The MICrONS program is predicated on the notion that it will be possible to achieve major breakthroughs in machine learning if we can construct synthetic systems that not only resemble the high-level blueprints of the brain, but also employ lower-level computing modules derived from the specific computations performed by cortical circuits.
Many contemporary theories of cortical computing suggest that the brain performs common sensory information processing tasks—such as detection and recognition of visual objects, sounds, and odors—with algorithms that progressively transform data through a series of operations, or “stages.” Each stage of processing is further theorized to occur within a discrete region of cortex. Although different theories suggest different mathematical bases for computation, it is commonly believed that neural algorithms employ data representations, transformations, and learning rules that are conserved across stages. It should therefore be possible to apprehend the neural computations underlying information processing (at least within a given sensory modality) by interrogating a small fraction of the entire cortex, so long as that fraction is judiciously selected to contain sufficient evidence of the representations, transformations, and learning rules of the algorithm(s) to which it contributes.
Neuroscience has a long history of inspiring innovation in machine learning, starting with the seminal work of McCulloch and Pitts in 1943. This influence is evident even in today’s state of the art “deep learning” systems, which are loosely modeled on hierarchical visual processing systems in the primate brain. However, the rate of effective knowledge transfer between neuroscience and machine learning has been slow because of divergent scientific priorities, funding sources, knowledge repositories, and lexicons. As a result, very few of the ideas about neural computing that have emerged over the past few decades have been incorporated into modern machine learning algorithms.
Previous attempts to foster collaboration between neuroscience and machine learning have been stymied in part by gaps in our knowledge about the brain. The majority of what is known about the brain today regards its operation at the “micro” scale (one or a few neurons) and the “macro” scale (hundreds of thousands or millions of neurons), and some of this information is indeed reflected in the design of leading artificial neural networks. In contrast, much less is known about the “mesoscale” cortical circuits (hundreds to tens of thousands of neurons) that implement the specific data representations, transformations, and learning rules of cortical information processing algorithms, and these are therefore absent from (or speculative in) existing machine learning solutions. It is likely that explicit knowledge and use of these computations is required to move beyond the current generation of “neurally-inspired” machine learning algorithms.
The MICrONS program aims to create novel machine learning algorithms that use neurally-inspired architectures and mathematical abstractions of the representations, transformations, and learning rules employed by the brain to achieve brain-like performance. To guide the construction of these algorithms, performers will conduct targeted neuroscience experiments that interrogate the operation of mesoscale cortical computing circuits, taking advantage of emerging tools for high-resolution structural and functional brain mapping. The program is designed to facilitate iterative refinement of algorithms based on a combination of practical, theoretical, and experimental outcomes: performers will use their experiences with the algorithms’ design and performance to reveal gaps in their understanding of cortical computation, and will collect specific neuroscience data to inform new algorithmic implementations that address these limitations. Ultimately, as performers incorporate these insights into successive versions of the machine learning algorithms, they will devise solutions that can perform complex information processing tasks aiming towards human-like proficiency.
MICrONS is organized in three phases, totaling five years in duration. During each phase, performers conduct targeted neuroanatomical and neurophysiological studies to inform their understanding of the cortical computations underlying sensory information processing and, concurrently, create neurally-derived machine learning algorithms that perform similar functions. Performers motivate their experimental and algorithmic designs by formulating and updating a conceptual model or “theoretical framework” for neural information processing in a given sensory modality. They use computational neural models (i.e., executable mathematic or algorithmic simulations of neurons and neural circuits) to establish a correspondence between the computations performed by biological wetware and the computations employed by their machine learning software. Each phase ends with an information processing challenge that assesses how well the new algorithms perform on increasingly challenging machine learning tasks: similarity discrimination in Phase 1, generalization and classification in Phase 2, and invariant recognition in Phase 3. Performers use the results of their experiments in each phase to guide their development of improved algorithms in the subsequent phase (in Phase 1, performers base their algorithms on the existing neuroscience literature).
The MICrONS
program comprises three Technical Areas (TAs). Although IARPA anticipates receiving a number
of holistic proposals responding to all three TAs, it recognizes that some
prospective offerors may have capabilities in only a subset of the overall
program scope, and wishes to maximize its opportunity to leverage these
capabilities. Therefore, offerors may
choose to propose to one, two, or all three TAs. Because achieving MICrONS program goals will
require significant collaboration across all three TAs, offerors who propose to
only one or two TAs should be prepared to work closely with performers in the remaining
TAs. The TAs in MICrONS are defined as
follows:
· TA1 – experimental design, theoretical neuroscience, computational neural modeling, machine learning, neurophysiological data collection, and data analysis;
· TA2 – neuroanatomical data collection; and
· TA3 – reconstruction of cortical circuits from neuroanatomical data and development of information technology systems to store, align, and access neural circuit reconstructions with the associated neurophysiological and neuroanatomical data.
Success in the MICrONS program will require extensive communication and cooperation between performers in all three TAs within or across teams. For example, in TA2, performers must collect neuroanatomical data about the same brain regions in the same brain specimens that are used in TA1 for neurophysiological studies; in TA3, performers must reconstruct neural circuits from the data collected in TA2; and in TA1, performers must analyze the neural circuits generated in TA3 and use the resulting insights in formulating their machine learning algorithms and theoretical frameworks. All offerors are therefore required to include in their proposal a detailed management plan and a detailed description of how their proposed technical approach in one or more TAs is likely to impact the other TAs.