    |
Scientific Interactions
| POGGIO - RIESENHUBER The computational framework sketched in the model (seen on homepage) is the conceptual tool to guide experiments and acts as the integrative "glue" across the different projects. Our Center will also be an experimental setup for testing the "meta" hypothesis -- relevant to several programs at NIH -- that models of neural systems can be powerful tools to understand complex brain functions and to drive interactive and synergetic interactions between different labs. Computational tools and theories already had a similar effect in physics and very recently in genomics. The hypothesis is that quantitative models of complex neural systems, when developed in close cooperation with experimental labs, can be tools to a) think about the problems (some cognitive problems are too complex for the qualitative, simple models used so far); b) make predictions, suggest and plan new experiments; c) analyze and interpret data; d) integrate experimental findings of different types and from different labs, drawing implications for future work from multiple sources of evidence. The computational component will rely on a close interaction between MIT (T. POGGIO) and Georgetown University (M. RIESENHUBER) and Caltech (C. KOCH). Tomaso Poggio was the Ph.D. advisor of C. Koch (in 1982) and they have continued to interact over the years. They have also been in close contact to develop the present project. They will interact electronically multiple times per week, and they will also meet frequently at MIT and/or Caltech. A component of the model, which is the key in our simulations to provide robustness to clutter and invariances to recognition, is the MAX operation, at various stages of the ventral stream. Thus, it is critical to evaluate whether there are sets of neurons along the ventral pathway that implement such an operation. The model predicts that the first stage showing a MAX-operation is a subset of cells (probably complex cells) in V1. Thus, FERSTER plans to carry out appropriate experiments, involving recordings in area 17 of the cat. On the biophysical side, Northwestern will focus on the circuitry prediction of the recognition model; initially, FERSTER will carry out intracellular recordings in cats at Northwestern University to test a key assumption of the recognition model -- that a MAX-like operation is performed at various stages in visual cortex. If the tests described in the Northwestern proposal will demonstrate that cells in area 17 do indeed compute a MAX-like function, FERSTER's group will next explore the underlying circuitry and interact with Caltech which will be simulating in detail, using compartmental models, cortical microcircuits. The intracellular recording technique will obviously be important at this stage of the project. So far, we have assumed an ideal MAX operation in the model and tested the robustness of the simulations for approximations of the ideal MAX. Once data from FERSTER's lab become available, RIESENHUBER and POGGIO will test the implications for properties of the recognition module of the experimental properties of the pooling operation. In particular, they will study invariance to position of the stimulus and the robustness to clutter, properties that depend critically on the type of pooling operation. The model simulations will interact closely with the planned experiments of DICARLO on selectivity, position invariance and robustness to clutter of IT cells during recognition. RIESENHUBER and POGGIO (in simulations) and DICARLO (in monkey psychophysics and physiology) will investigate the neural mechanisms underlying recognition in clutter by investigating the effect of introducing an additional object into an IT neuron's receptive field. As we mentioned, KOCH will perform psychophysical experiments in humans to determine the effects of clutter on a task's attentional requirements. Based on these experimentally obtained constraints, the computational model of saliency previously developed in KOCH's group will be integrated with the feed-forward recognition model of POGGIO and RIESENHUBER to implement attentional modulation of object recognition. In this work KOCH will interact closely not only with RIESENHUBER and POGGIO along computational lines to extend the recognition model to account for attentional effects, but also directly with DICARLO, through KOCH's computational efforts. Both will be testing the effect of clutter and the increasing difficulty in recognition, one at the single cell level in monkeys and the other at the behavioral level in humans. One of
the predictions of the global architecture of the existing model of recognition
is that both categorization and identification are supported by the same
computational mechanisms. MILLER will test this prediction by recording
in inferotemporal and prefrontal cortex in the macaque monkey. MILLER's
experiments will interact with the other investigators mainly through
the model. Data on identification and categorization in IT will, via the
model, affect the interpretation of the results of DICARLO on sensitivity
to clutter in IT and vice versa. MILLER and DICARLO's projects have also
a direct synergy. In the experiments of DICARLO, monkeys will be trained
with a small number of target objects. The expectation is that some neurons
will become tuned to them, as it was the case in Logothetis experiments
[87] (and also DICARLO and Maunsell [31]). In MILLER's experiments, training
will be done with a whole class of objects, using samples from a continuous
morph space. It will be instructive for our understanding of object class
representation in cortex to compare the resulting neural representations
obtained in the two labs using the model to check their consistency with
each other and with existing data. |