For the second year in a row, the Machine Learn­ing Week Europe con­fer­ence has been held online. Unfor­tu­nately, that meant no ming­ling in the cor­ridors, flag­ging down waiters with trays of hors d’oeuvres, and clink­ing glasses with fel­low ML enthu­si­asts and prac­ti­tion­ers. For­tu­nately, how­ever, the con­fer­ence did suc­ceed in bring­ing together some fant­astic tal­ent for some world-class talks.

Here’s a quick sum­mary of some that stood out for me.

30 Golden Rules of Deep Learn­ing Performance

Siddha Ganju – NVIDIA

Siddha provided an extremely access­ible and com­pre­hens­ive list of speed-ups for mod­ern Deep Learn­ing pro­jects (with a bit of cheeky humour sprinkled in). The talk comes on the heels of her new book Prac­tical Deep Learn­ing for Cloud, Mobile & Edge. Also very cool: everything from the talk is avail­able if you send an e‑mail to PracticalDLBook@gmail.com with the title “wget School of AI presentation”.

Explain­able AI in Deep Brain Medicine

Afsaneh Asaei & Ozgur Polat – UnternehmerTUM and the Schoen Clinic of Neurology

In a phe­nom­en­ally cre­at­ive use of Machine Learn­ing, the voice pat­terns of Parkinson’s patients can be used to identify if their con­di­tion is “flar­ing up” (enter­ing an “off” period). Audio record­ings are changed into spec­tro­gram images, which are then fed into a neural net­work (1 CNN layer, 2 RNN lay­ers, 1 FC layer). The net­work out­puts a bin­ary value cor­res­pond­ing to the state of the patient’s con­di­tion. The full art­icle is avail­able at Jain et al. 2021.

Data Sci­ence at The New York Times

Chris Wig­gins – New York Times

When you think “New York Times” the men­tal imagery is likely unfurl­ing a can­vas of crisp typeface onto your din­ing room table, cof­fee in hand. You prob­ably do not think of innov­at­ive data sci­ence pro­jects. The real­ity is that the New York Times oper­ates a large, and won­der­fully cre­at­ive, data sci­ence depart­ment, headed by Chris Wig­gins. Using a wealth of data gen­er­ated by online read­er­ship and sub­scrip­tions, the team plays – and “plays” is really the cor­rect word here – with dif­fer­ent ima­gin­at­ive projects.

As an example, the team once thought to them­selves: “I won­der if we could fig­ure out how art­icles make people feel.” Through ques­tion­naires for online read­ers, they were able to cre­ate an NLP + clus­ter­ing sys­tem able to identify, for a given art­icle, emo­tions likely to be triggered within read­ers. Through this, advert­isers could opt to show their ads along­side art­icles likely to invoke a spe­cific emo­tion. The res­ult was a huge suc­cess, with bet­ter advert­ise­ment rev­enue, and a more focused under­stand­ing of the impact of pub­lished stories.

A sum­mary of this spe­cific pro­ject can be found here.

Pre­dict­ive Maintenance

Mar­cus Gross – INWT Statistics

What is the best way to approach a pre­dict­ive main­ten­ance use case? Although there is no one “right” way of doing things, Mar­cus sug­gests a few gen­eral tips from recent research in the domain:

  • Recently, there has been some research into apply­ing Cox Pro­por­tional Haz­ard to Deep Learn­ing. This can be a good solu­tion, espe­cially with very sparce data­sets con­tain­ing few instances of failure/breakdown. A worked example can be found here.
  • To under­stand which fea­tures (or fea­ture com­bin­a­tions) are the strongest indic­at­ors of fail­ure, use Shap­ley val­ues, or (more spe­cific­ally) SHAP val­ues.
  • Weibull Dis­tri­bu­tion can be a use­ful met­ric for help­ing the industry decide when to ser­vice machinery. A worked example in Keras can be found here.
  • If you have a wealth of data with many examples of fail­ure, then a simple answer may work best: a cus­tom search algorithm that iden­ti­fies parts or fea­tures that con­trib­ute to high rel­at­ive failure.
  • Finally, XGBoost is the weapon of choice when train­ing tree mod­els pre­dict­ing fail­ure. Don’t skip it!

Lever­aging Unstruc­tured Data in Insurance

Ray­mond von Es – Milliman

Does your com­pany have a huge whack of cus­tomer call record­ings? How do you extract value?

Some points from an expert:

  • It is imper­at­ive that cus­tomer calls are recor­ded on 2 audio chan­nels (1 for cus­tom­ers, 1 for call centre employ­ees). Without doing this, it can become a huge head­ache to split them later.
  • GCP speech-to-text API (batch pro­cessing) works great
  • Tech­nical or domain-spe­cific words often have to be treated separately
  • On Ger­man data GCP is good enough for Sen­ti­ment Ana­lysis & Topic Identification
  • On Eng­lish data it is also some­times pos­sible to cre­ate auto­mated summaries

The Role of Tensor­Flow, PyT­orch, and Friends in a Med­ical Device – the Reg­u­lat­ory Perspective

Oliver Haase – Val­id­ate ML

If I’m look­ing to get an AI sys­tem cer­ti­fied by TÜV, can I used ML pack­ages and lib­rar­ies, or should everything be coded “from scratch”?

Oliver Haase, of Val­id­ate ML, provides the answer to this ques­tion: it will depend on the industry.

In gen­eral plant deploy­ment, pack­ages can be used without worry. Cer­ti­fic­a­tion will prin­cip­ally depend on factors such as (among others):

  • Human oper­at­ors being present to over­see the AI system
  • The AI sys­tem being stop­pable at any given moment
  • The AI sys­tem being impen­et­rable to hack­ers (air-gapped)
  • Explain­ab­il­ity

How­ever, for med­ical devices, where human lives may be at stake, the require­ments are stricter. Pack­ages do need to be val­id­ated before use. For an example of why this is neces­sary, con­sider this situ­ation, in which a bug in the data aug­ment­a­tion method com­bin­ing NumPy with PyT­orch leads to unreal­istic accur­acy estimates.

But the pro­cess of val­id­a­tion need not be cum­ber­some! If an effi­cient val­id­a­tion pipeline is cre­ated, many steps can be reused. For the expli­cit reg­u­la­tions, check ISO 13485 and IEC 62304.

Dis­cov­er­ing Key Top­ics from Real-World Med­ical Inquir­ies via Nat­ural Lan­guage Pro­cessing at Bayer

Angelo Ziletti – Bayer

Phar­ma­ceut­ical com­pan­ies receive count­less med­ical inquir­ies, often in the form of short text blurbs. So how do you tackle them? Ziletti’s team used vari­ous NLP and clus­ter­ing meth­ods to sort and group by top­ics. The task was com­plic­ated by the fact that inquir­ies would some­times be a single sen­tence or only a few words.

The res­ult­ing col­lage of top­ics and patient con­cerns provided an action­able roadmap for which areas should be focused on in patient com­mu­nic­a­tion, as well as for gen­eral improve­ments in the med­ical con­text. The full paper is avail­able here.