Weekly Research Meetings

CBMM Weekly Research Meeting: Tales from robotics: teaching iCub to recognize objects

Oct 7, 2014 - 4:00 pm
IIT iCub
Venue:  Harvard University: Northwest Bldg, Room 243 Address:  52 Oxford Street, Harvard University Northwest Building, Cambridge, 02138 Speaker/s:  Dr. Carlo Ciliberto, LCSL MIT

CBMM Thrust 5 – Theories for Intelligence
For more information regarding the iCub, please visit  http://www.icub.org/

Organizer:  Tomaso Poggio Lorenzo Rosasco

CBMM Weekly Research Meeting: Atoms of recognition in human and computer vision

Sep 16, 2014 - 4:00 pm
Prof. Shimon Ullman
Venue:  Harvard University: Northwest Bldg, Room 243 Address:  52 Oxford Street, Harvard University Northwest Building, Cambridge, 02138 Speaker/s:  Prof. Shimon Ullman

Abstract:

The human visual system makes highly effective use of limited information: it can recognize not only objects, but severely reduced sub-configurations in terms of size or resolution. Minimal images are useful for the interpretation of complex scenes but they are also challenging because by their nature they are non-redundant stimuli. Human studies and computer simulations show that humans and existing models are very different in their ability to interpret minimal images. I will discuss possible implications to the representations used for recognition, brain mechanisms, and algorithms for the interpretation of complex scenes.

CBMM Weekly Research Meeting

Sep 9, 2014 - 4:00 pm
CBMM
Venue:  Picower Seminar Room (46-3310) Address:  43 Vassar Street, MIT Bldg 46, Room 46-3310, Cambridge, 02139

Meeting Notes:

Admin
● New (?) members
● 25 postdocs (affiliated or participating)
○ announce in labs/PIs
● Mailing List
● Separate grad student list (use both lists for postdoc meetings)
● Meetings: 1 meeting every 3 weeks (MIT or Harvard)

Topics
● Organization for this year (Content)
● Role of CBMM postdoc group:
○ Independent postdoc activities
○ Arrange regular Center meetings (?)
○ Summer Course
● Summer Course (recap + plan)

Content of Meetings
Content of meetings
● Regular technical talks (resume with everyone that has not presented)
○ Random order or volunteers
● Training sessions (by anyone that wants to volunteer on a subject)
● External speakers (e.g. postdocs or researchers from other institutes, CSAIL etc.).
● Seek other meetings with industrial partners
● (Open to graduate students)

CBMM Summer Course 2015
Meeting (dedicated)
● Outline tentative schedule for 2015 SS
● Thursday 9/18 at 4pm ? (leyla lisik@mit.edu)
● Official TAs, Grads, Postdocs etc.

People
● Double number of official TAs (no unofficial TAs)
● (perhaps) Increase # official students to 30 or 35 (if we can advise well)
● All speakers should commit to spend at least 2 days with scheduled time to interact with
students (lunch and dinner)
● There should be one person per each of the 5 thrusts at all times

During
● Computer room in the third floor | Lectures could be in the basement or in the auditorium we
used for evening lectures.
● The first day of the school should reflect the overall balance of the school not only its
computational aspects
● Ideally we should have overviews in the morning and hands-on activities in the afternoons
● Students should get feedback on projects from faculty and TAs during the summer course period
not only at the end
● There should be enough discussion after each project presentation. Faculty should be involved.

Projects
● When we advertise applications we should make a condition that accepted students will have to
take an online MATLAB tutorial before the summer course
● We should describe possible projects well before the summer course takes place
● We should make clear at the beginning of the school what the rules are for projects: group
projects are fine but individual projects are fine too

Various
● IIT is planning to bring the iCub for the 3 weeks and one person who can help with projects
● Wearables like Google Glass and software to record and analyze (Astar)
● Eye tracker in the wild and a person to support its use (Nancy)

CBMM Weekly Research Meeting: Visual-object processing culminates in inferior temporal cortex (IT)

May 23, 2014 - 4:00 pm
Visual-object processing culminates in inferior temporal cortex (IT)
Venue:  MIT: McGovern Institute Reading Room, 46-5165 Address:  43 Vassar Street, MIT Bldg 46 , Cambridge, MA 02139 United States Speaker/s:  Prof. Bevil Conway, Wellesley

Topic:  Progress on CBMM Challenge

Abstract:

To assess the organization of IT, we measured functional magnetic resonance imaging responses in alert monkeys to achromatic images (faces, fruit, bodies and places) and colored gratings. IT contained multiple color-biased regions, which were typically ventral to face patches and yoked to them, spaced regularly at four locations predicted by known anatomy. Color and face selectivity increased for more anterior regions, indicative of a broad hierarchical arrangement. Responses to non-face shapes were found across IT, but were stronger outside color-biased regions and face patches, consistent with multiple parallel streams. IT also contained multiple coarse eccentricity maps: face patches overlapped central representations, color-biased regions spanned mid-peripheral representations and place-biased regions overlapped peripheral representations. These results show that IT comprises parallel, multi-stage processing networks subject to one organizing principle.

CBMM Weekly Research Meeting: Parsing Objects and Scenes in Two- and Three-Dimensions

May 16, 2014 - 4:00 pm
Parsing Objects and Scenes in Two- and Three-Dimensions
Venue:  MIT: McGovern Institute Singleton Auditorium, 46-3002 Address:  43 Vassar Street, MIT Bldg 46, Cambridge, 02139 United States Speaker/s:  Alan L. Yuille, Professor & Investigator, UCLA

Topic: Progress on CBMM Challenge

Abstract:

We continue the series of weekly discussions and reports on each CBMM challenge question describing progress and problems of ongoing work at CBMM.

Thrust 5 is focused on models for the CBMM challenge that can answer CBMM challenge questions while being consistent with human behavior and neural data. This talk presents three recent studies on detecting and parsing objects and scenes and discusses how they contribute to the CBMM challenge. We first address the “what?” problem of detecting animals and animal parts (in a newly labelled dataset) and show the advantages of part-sharing (X. Chen et al. 2014). Next, within the same “what?” problem, we describe an approach to parse humans and estimate their  three-dimensional structure from single images (C. Chen et al. CVPR 2014).  Finally, we describe “psychophysics in the wild” for rapid detection of objects in complex scenes in a newly labelled dataset (Y. Li et al, CVPR 2014). We conclude discussing how these approaches should be extended to meet the CBMM challenge and other efforts at CBMM.

CBMM Weekly Research Meeting: Ongoing fMRI Investigations of Social Perception in the STS

May 9, 2014 - 4:00 pm
Venue:  MIT: McGovern Institute Reading Room, 46-5165 Address:  43 Vassar Street, MIT Bldg 46 , Cambridge, MA 02139 United States Speaker/s:  Ben Deen

Collaborator: Nancy Kanwisher, Rebecca Saxe; CBMM Thrust 4 – Social Intelligence

Progress of CBMM Challenge – Social Intelligence

Abstract:

The central aim of Thrust 4 is to understand nonverbal social perception, or the ability to make high-level social inferences from perceptual information, in order to allow the development of models able to answer CBMM challenge questions about people and their social interactions from images and videos. The study of social perception is in its relative infancy: we have a catalog of intriguing behavioral findings on what sorts of inferences can be made, but don’t have a sense of the full scope of these abilities, nor any understanding of their computational basis or neural implementation.

Here, we propose to investigate the functional organization of the neural machinery for social perception as a first step toward understanding the computational and neural architecture of these abilities.  Specifically, we target the superior temporal sulcus, which has been argued to play a role in a large number of social perceptual and cognitive processes: the perception of faces, biological motion, and voices, as well as the ability to understand language and mental state content.

In experiment 1, we use fMRI to measure STS responses during a number of different social perceptual tasks, and compare responses within individual participants.  We find that the STS has a strong spatial organization, with distinct processes eliciting distinct patterns of activity.  Furthermore, we find both evidence for selective STS subregions, which respond strongly to one task and not others, and multifunctional regions in which responses to multiple tasks overlap.

In experiment 2 we further investigate a particularly strong case of overlap from experiment 1 – overlap between responses to dynamic faces and vocal sounds. […]

CBMM Weekly Research Meeting: Progress of CBMM Challenge – Development of Intelligence

May 2, 2014 - 4:00 pm
Venue:  Harvard University: Northwest Bldg, Room 243 Address:  52 Oxford Street, Harvard University Northwest Building, Cambridge, 02138 Speaker/s:  Noah Goodman; CBMM Thrust 1 – Development of Intelligence

Progress on the CBMM Challenge Questions: What is there? and Who is there?

Abstract:

Thrust 1 presents this week episode in our series of weekly discussions on progress on the CBMM challenge questions. The thrust is focused on the development of intelligence and on how to model it. It is therefore appropriate to step back and ask: what role do compositionality, probabilistic inference, and intuitive theories have to play in our understanding of understanding? Noah Goodman will start by discussing the Probabilistic Language of Thought hypothesis, including taking a quick tour of the web book probmods.org and the models repository forested.org. Noah will then ask how language interfaces with non-linguistic understanding, and how context mediates this interaction, sketching an architecture for natural language semantics and pragmatics, grounded in probabilistic intuitive theories. Noah plans to extend the basic architecture to figurative language, such as hyperbole, and argue that compositionality of thought it central but deeply buried in everyday language. This will be the motivation to ask what formal semantics and pragmatic effects we may need to engage with even for the (mostly non-linguistic) CBMM challenge.

CBMM Weekly Research Meeting – Top-down Control of Attention

Apr 25, 2014 - 4:00 pm
Top-down Control of Attention
Venue:  MIT: McGovern Institute Seminar Room, 46-3189 Address:  43 Vassar Street, MIT Bldg 46, Cambridge, MA 02139 United States Speaker/s:  Robert Desimone; Thrust 2 – Circuits for Intelligence and Thrust 5 – Theories of Intelligence

Progress on the CBMM challenge questions: What/Who is there?

Abstract:

We continue the series of weekly discussions and reports on each CBMM challenge question describing progress and problems of ongoing work at CBMM.

Thrust 2 is focused on the Neural Circuits and on how models for the CBMM challenge should be consistent with the brain neural data. One of the projects in the thrust focuses on the neural correlates of visual attention, which plays a key role in the answer to most of the CBMM open set of questions, probably in the form of task-dependent visual routines. I will describe very recent results about the neural correlates of so-called object attention. The voluntary control of visual attention to behaviorally relevant stimuli is thought to involve “top-down” feedback to visual processing areas. For spatially-directed attention, one key source of top-down attention is the frontal eye fields (FEF). My lab found that feedback to visual cortex from FEF causes enhanced responses to stimuli at attended locations, and leads to synchronized neural activity in the gamma frequency range between FEF and visual processing areas. Recent evidence suggests that the pulvinar may also serve as an important relay of attentional feedback to visual cortex, and it may also serve to desynchronize cortical activity in the alpha frequency range. Many aspects of the neural response changes with attention are explained by top-down inputs to a simple cortical circuit containing excitatory and inhibitory neurons. The neural basis of feature, or object attention has been much more difficult to understand.

One possibility is that attention to objects with particular features causes spatially directed attention to be directed to those objects, utilizing known pathways for spatial attention. Another possibility is that attention to objects or features such as faces, colors, or shapes, depends on feedback to visual cells that are selective for those features, biasing activity in favor of those stimuli. Such a mechanism would be similar to what is thought to mediate visual recall memory. We have recently found evidence for both types of mechanisms in prefrontal cortex of humans and monkeys.

CBMM Weekly Research Meeting: Progress of CBMM Challenge – Theories of Intelligence

Apr 18, 2014 - 4:00 pm
Venue:  Harvard University: Northwest Bldg, Room 243 Address:  52 Oxford Street, Harvard University Northwest Building, Cambridge, 02138 Speaker/s:  Winrich Freiwald and Joel Z. Leibo; CBMM Thrust 5 – Theories for Intelligence

Title: On the neural mechanisms of face recognition: from experiments to theory

Abstract:

Object recognition, the ability to identify an object despite vast changes in appearance due to changes in lighting or orientation, is a major accomplishment of the primate brain, as a result of which we can recognize objects with an ease belying the daunting complexity of the computational challenges involved. Understanding not only the algorithms used to answer the questions What is there? Who is there?, but also the underlying neural mechanisms  is a major problem for both the experimental and the theoretical neurosciences – and one of the main items in the CBMM challenge. To decipher the mechanisms of object recognition, we have taken advantage of a unique model system evolution has provided us with. The temporal lobes of macaque monkeys, like those of humans, contain neural machinery to support face recognition. The machinery consists of a fixed number of discrete patches of face-selective cortex that can be localized with functional magnetic brain resonance imaging. The three main organizing features of this system, concentration of cells encoding the same complex object category into modules, spatial separation of modules with different functional roles for face processing, and integration of modules into an interconnected network, make it possible to break down the process of face recognition into its components. In this talk, we will present the experimental data we have obtained over the last years that characterize the major properties of this system. In particular, we will discuss results showing that a mirror symmetric face representation is computed as an intermediate step between view-tuned and view-tolerant representations. In the second part of the talk we will describe how this finding was influential in the early conception and development of M-theory, currently one of the research directions in the CBMM. We will describe and discuss a computational model of the face-processing system that is derived from M-theory and predicts and explains the core experimental finding.  This interdisciplinary project on the neural mechanisms for the processing of a stimulus set of utmost relevance for social cognition, bridges activities between three Thrusts of the CBMM.

CBMM Weekly Research Meeting: What can mice tell us about visual intelligence?

Apr 11, 2014 - 4:00 pm
Alignment example
Venue:  MIT: McGovern Institute Reading Room, 46-5165 Address:  43 Vassar Street, MIT Bldg 46 , Cambridge, MA 02139 United States Speaker/s:  Dr. Michael Buice, Allen Institute for Brain Science, CBMM Thrust 2: Circuits for Intelligence and Thrust 5: Theories for Intelligence

Progress on the CBMM challenge questions: What/Who is there?

Abstract:
In the Jeopardy/Watson effort every week on Friday there was an evaluation of performance. We continue the series of weekly discussions and reports on each CBMM challenge question [e.g. What is there? What will happen next? What are they doing? etc.] describing progress and problems of ongoing work at CBMM.

The Allen Institute for Brain Science uses a big science approach to neuroscience, integrating experiment, modeling and theory. One of the Institute’s projects, MindScope, is an exploration of the mouse visual system from a multitude of coordinated approaches. I will give an overview of the MindScope project and the “C3” approach – Components, Computation, and Cognition, and discuss how the goals of MindScope relate to the goals of CBMM. Important goals along our Cognition axis are Object Recognition, Attention, and Decision Making in the mouse cortico-thalamic system. The M-Theory framework provides a feedforward architecture which computes invariant representations of the visual field, facilitating fast object recognition which mimics human performance on recognition tasks within about 100ms. Several results suggest that recognition beyond 100ms or so involves feedback. I will discuss some preliminary ideas on extending the M-Theory framework to include feedback, in particular by incorporating elements of probabilistic inference into the model.

Pages

Subscribe to Weekly Research Meetings