Home > Meeting Topics and Material > Weekly QUEST Discussion Topics and News, May 31st

Weekly QUEST Discussion Topics and News, May 31st

We defined a measure of intelligence associated with a task to be the performance function of the agent (or set of agents) for that task (example finding cancer in a set of mammograms where you have defined a particular performance function like the sensitivity -> probability of detecting cancer when it is present in the woman)

We went to great lengths in the WCCI material to define performance functions / tasks as critical prior to any discussion on intelligence or intelligence amplification

What I am working on this week is an approach where we can emphasize the critical importance of ‘alignment’ to maximizing IA – the thought is to do that I have to define a task and as part of the performance function I have to define the manner in which the human and computer are ‘fused’ to provide an answer in the arm that has both – and how they can function independently in arms where they are not combined – that will give us intelligence estimates for what we called ‘h’ and what we called ‘c’ in the ‘ari gesher’ blog discussion

I can have a particular human do a set of mammograms and get an estimate of the performance – f(t) = h, I can have an algorithm like Second Look execute on that same set of mammograms and get an estimate of the performance –f(t) = c,

Using the notation of the Ari blog these are the a(h) and the a(c) –

Now to get an estimate of a(h,c) I have to constrain the manner the two agents combine their analysis to make a decision – one of my patents – 6,115,488 defines very specifically how to combine detections to maximize sensitivity – and for this discussion that is the performance function – f, in that patent all the detections of the human alone survive the ‘fusion’ – (thus maximizing the sensitivity and also acknowledging the purpose of CAD is NOT to prevent ‘work-ups’ of regions that concern doctors but the purpose of CAD is to help docs find regions that may have missed but are detectable by a machine CAD) – so the human looks at the output of the computer CAD and decides whether he/she wants to add additional locations of concern to their original list of regions based upon the CAD detections – NOTE in this fusion the performance function can NOT be reduced by the addition of the computer aide. Thus the overall intelligence of the human computer combination as measured by the sensitivity performance function is greater than or equal to the human alone intelligence! So in the Ari blog notation a(h,c) >= a(h).

So my query this week for QUEST is how can we make statements associated with friction / alignment and its impact on cancers that were overlooked by the human that upon re-examination based on the fact that the CAD marked the location will maximize IA. In the fusion rule I’ve patented above the answer seems obvious to me. If the CAD marks something that the human upon re-examination can NOT align to (it doesn’t look like cancer) then the human will not add it to the list of regions from their original review and thus it can NOT increase the sensitivity.

Note: we can argue that the maximum IA would be if the computer aid provides answers that the human can’t decipher using human perceptual powers it just requires in a ‘fusion’ rule that is different. Such combinations may in fact NOT want maximal alignment with human cognition – but they have to be trusted by the human as a result of performance – that was the point of the Prof Kurzweil position in our conversation with him last week.

Weekly QUEST Discussion Topics and News May 31

  1. No comments yet.
  1. No trackbacks yet.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: