Wednesday, April 23, 2008

Dispatch from CNS: The TB East Perspective

The CNS meeting in San Francisco, as usual, was a smorgasbord of cognitive and neuro science, and, as usual, ranged from the ridiculous to the sublime. I will refrain from identifying what I considered ridiculous (buy me a beer and I'll point to a bunch of stuff) and also what I considered sublime (what the hell do I know?), but there were a bunch of interesting things. Here's a smorgas-reaction to the smorgas-bord.
  • Lolly Tyler organized a symposium on visual object recognition on the day before CNS. This was my favorite part of the meeting. I anticipated a bunch of discussion about the role of top-down mechanisms and prediction; interestingly, the main message I got from this pre-symposium symposium was that the feed-forward mechanisms of the visual system (in particular, the ventral stream) get you really far in terms of object recognition subroutines. Poggio from MIT presented the computational model he's developed over the years and argued that feed-forward mechanisms comfortably account for the first 100 ms of perception of single images. That is not to say that there aren't important contributions of 'analysis by synthesis' operations, but the early aspects of this are well accounted for by feed-forward aspects of the visual system. Shimon Ullman from the Weizman Institute presented an object recognition model in which she appealed to a hierarchical organization of features of an object. The challenge, obviously, is to figure out what the features are: they should be less generic than, say, Gabor filters but more generic than the specific object. Obviously. So what is the right 'alphabet'? The challenge is to find the features with the largest amount of information. Great talk. I'd like to adapt some of these ideas to hearing and speech. Moshe Bar from Harvard gave a very interesting talk about how a quick glance at low-resolution, low spatial frequency information can be used to generate predictions about possible objects that are then 'verified' by the ventral stream, high-resolution, high spatial frequency stuff. Bar has a TICS paper that describes his model, an interesting read. DiCarlo from MIT presented neurophysiological data from recordings in the ventral stream. The most provocative part of his stuff argued that IT neurons reflect invariant properties of target objects but V4 neurons do not. I'd like to believe that, but what is the magic mojo from which derives that mythical invariance? Anyway, very cool talk addressing a very deep problem. There were some other presentations that were not quite as compelling.
  • On Tuesday morning, there was a symposium on the anterior temporal lobes and semantic memory. This topic has been discussed in this blog at length, and one of the presentations (by Richard Wise) was amusingly summarized today by the blogger Neurocritic (see Greg's earlier post). In that symposium, Matt Lambon Ralph gave a nice presentation of his research, which has been discussed in this blog (see earlier postings). Wise's talk was less narrowly focused on ATL and semantic memory/semantic dementia, but rather constituted an amusing and sharply formulated tour through neurolinguistic history. Richard highlighted a few papers that he clearly values but that he cleverly and humorously criticized. Richard is one of the people in the field who is the most rigorous about keeping us honest about the anatomy. And he's quite right to do so. And in terms of good publicity for Hickok & Poeppel 2007, he did mention our ideas articulated there as not entirely useless. Richard is thinking very deeply about what it can or must mean to have modality-independent language processing, and he has done a series of experiments investigating this. This is difficult stuff, and I think this kind of research is making important incremental progress in our understanding of the 'homework problem' of figuring out the anatomic aspects of the functional architecture.
  • Other stuff I found noteworthy: The dinner with Sonja Kotz, Jonas Obleser, Matt Lambon Ralph, Richard Wise, and others at Cosmopolitan. Great soup. Nice dessert, too. Weird pasta. I'd never met Manuel Carreiras, and one evening I had drinks with him and my friend Päivi Helenius. These are cool people to hang out with. Smart as hell and non-nerdy.
  • There were lots of interesting posters, and lots of utter schlock. Some of the stuff was like painfully stupid. Stuff I rather liked included the growing body of evidence on lexical decomposition in a variety of contexts. There were terrific posters on this by the Carreiras and colleagues, Marantz, our own shop (he says humbly), and other labs as well. It's pretty clear that there's 'total convergence' on the view that there is extensive decomposition during lexical access and language processing more generally. The cool thing is that we're now figuring out the detailed mechanisms and time course and worrying less about the ideological mumbo-jumbo. One more thing for good measure: I was quite fascinated by the growing body of work on infants using NIRS. There was cool work on this by Heather Bortfeld from Texas A & M and by Silke Telkemeyer from Berlin. This work is still very much in its infancy, haha, but once a more normative set of data have been acquired for various aspects of perception, NIRS can become an exciting view on the earliest possible chunks of life. For example, the Berlin group presented data from three-day-old babies from which they recorded simultaneous ERP and NIRS.

No comments: