NPPS Leadership Team

[Torre Wenaus <wenaus AT gmail.com>]

Hi,
On the brighter side, I talked to Hong and Dmitri about this today and they emphasize that we need our own HEP/NP plan and vision, we should be consistent with lab/CSI level plans but not dependent on it or deferring to it. My paraphrase. They support the EIC and HEP content in the doc, and the immediate objectives of building community around the many activities already underway, and leveling up AI/ML knowledge. They want me to organize and I told them I'd like to use this informal group to plan what we do, e.g. a ~monthly meeting to share what we're doing and get better educated on AI/ML? We've spent almost two years in the "CSI & Co will build a lab plan and vision" mode, that phase is thankfully ended, and thank you to the perceptive PEMP reviewers for jogging our management into action. Let's set our own plans and direction, making use of our CSI colleagues insofar as they are useful, and they can be useful in direct collaboration e.g. as Brett has described to me regarding his collaborative project/LDRD with them. One concrete thing is the plan Brett has already been working out with Meifeng at CSI to have a series of 'leveling up' meetings between CSI and Physics.
  Torre

On Tue, Jan 26, 2021 at 11:47 AM Anže Slosar <anze AT bnl.gov> wrote:

Hi all,

So let me add to this thread that regardless of the fact that they were
given as boundary conditions, I share many of Brett's concerns.

off the recrod: after working with supposed "AI/ML" experts on the
SciDac project for the past 4 years I am underwhelmed. The two main
issues are:

- they do not seem to be cutting edge in the sense google and facebook
  are. What they do in practice is to play with the box of lego bricks
  provided by the industry giants who seem to have the real brains
  working on this. But they are not either capable or willing to design
  special bricks that we would need.

- because giants are interested recognizing cats and dogs rather than
  careful propagation of noise, we end up cats and dogs. The real
  problem is that it is easy to train network to produce something that
  is much more impressively correct by eye than a simple old school
  method. But it is very difficult to ascertain noise properties and
  subtle biases from imperfect training datasets. Which makes these
  methods often borderline useless.

We need to find a constructive way of engaging with these people and
educating them about our needs rather than letting them apply
variational encoders and GANs to everything that moves.

        Anže


> Date: 2021-01-25T16:26:16-0500
> From: Brett Viren <bv AT bnl.gov>
> Subject: Re: [Phys-npps-mgmt-l] One slide presentation at the next Monday 4pm meeting
>
> Hi Elke,
>
> Aschenauer Elke-Caroline <elke AT bnl.gov> writes:
>
>> I’m must admit lost with your comments are you referring to what Torre added
>> on HEP or are you unhappy about the entire document or something else.
>
> Oh heck!  Okay, well I feel super chagrinned now.  I thought there are
> only two pages in the document so my knee jerk was only about the first
> "vision".  Sorry for my overreaction!
>
> After actually scrolling down, the EIC and HEP visions are much more
> lined up with my hopes for AI/ML at BNL.
>
> -Brett.

--

-- Torre Wenaus, BNL NPPS Group, ATLAS Experiment

-- BNL 510A 1-222 | 631-681-7892 |  wenaus AT gmail.com | npps.bnl.gov | wenaus.com

-- NPPS Mattermost room: https://chat.sdcc.bnl.gov/npps/channels/town-square