Thursday, January 26, 2017

Deep learning for cancer research

A recent interesting news from Stanford regarding identification of skin cancer using deep leaning for images.

A different project featured by NVIDIA is using deep learning for breast cancer research, where they claim that the error went down 85%.

Unrelated, I heard today about Grail who raised 100M$ for cancer detection in blood tests. Grail raised money from Amazon, Google, and Microsoft (Bill Gates).  Looking at their career page they are also looking for deep learning researchers.

Another interesting company is Zebra Medical Research which shares medical data with researchers in return for a fraction of future revenues.

Following this blog post publication, my friend Assaf Araki from Intel sent me a reminder for Intel's cloud cancer research initiative. Broad MIT institute joined last year.

Wednesday, January 25, 2017

Amazing deep learning visualization

I found this amazing deep learning visualization:
The tool is written by Daniel Smilkov and Shan Carter from Google Brain's team.

It is a great tool to understand using small examples the network operation.

The tool promised:

It took me 5 minutes to find a configuration which breaks it ! :-)

Here it is:

Note that both Test loss and training loss are NaN. I am waiting for the fix. (I see this error was already reported)

Thursday, January 19, 2017

TensorFlow to support Keras API

I found this interesting blog post by Rachel Thomas. My favorite quote:

Using TensorFlow makes me feel like I’m not smart enough to use TensorFlow; whereas using Keras makes me feel like neural networks are easier than I realized. This is because TensorFlow’s API is verbose and confusing, and because Keras has the most thoughtfully designed, expressive API I’ve ever experienced. I was too embarrassed to publicly criticize TensorFlow after my first few frustrating interactions with it. It felt so clunky and unnatural, but surely this was my failing. However, Keras and Theano confirm my suspicions that tensors and neural networks don’t have to be so painful. - production environment to serve TensorFlow models

I recently stumbled upon - an open source production environment to serve TensorFlow deep learning models. By looking into Giuhub activity plots I see the Chris Fregly is the main force behind it. is trying to solve the major headache around scoring and maintaining ML models in production.

Here us their general architecture diagram:

Here is a talk by Chris: 

Alternative related systems are, (sold to SalesForce), (sold to Cloudera), Domino Data Labs and probably some others I forgot :-)

BTW Chris will be giving a talk at AI by the bay conference (March 6-8 in San Francisco). The conference looks pretty interesting. 

And here is a note I got from Chris following my initial blog post:

Thanks for the mention, Danny! Love your work.

Here's an updated video:

Here's the jupyter notebook that powers the entire demo: 

I asked Chris which streaming applications he has in mind and this is what I got:

We've got a number of streaming-related Github issues (features) in the works: here are the some relevant projects that are in the works: - working with the Subscriber-Growth Team @ Netflix to replace their existing multi-armed bandit, Spark-Streaming-based data pipeline to select the best model to increase signups. we're using Kafka + Kafka Streams + Spark + Cassandra (they love Cassandra!) + Jupyter/Zeppelin Notebooks in both Python/Scala. - working with the Platform Team @ Twilio to quickly detect application logs that potentially violate Privacy Policies. this is already an issue outside the US, but quickly becoming an issue here in the US. we're using Kafka + custom Kafka Input Readers for Tensorflow + Tensorflow to train the models (batch) and score every log line (real-time). - working with a super-large Oil & Gas company out of Houston/Oslo (stupid NDA's) to continuously train, deploy, and compare scikit-learn and Spark ML models on live data in parallel - all from a Jupyter notebook. - working with PagerDuty to predict potential outages based on their new "Event" stream which includes code deploys, configuration changes, etc. we're using Kafka + the new Spark 2.0 Structure Streaming.  

What are the main benefits of vs. other systems? - the overall goal, as you can probably figure out, is to give data scientists the "freedom and responsibility" (hello, Netflix Culture Deck!) to iterate quickly without depending on production engineers or an ops group. - this is a life style that i really embraced while at Netflix. with proper tooling, anyone (devs, data scientists, etc) should be able to deploy, scale, and rollback their own code or model artifacts. - we're providing the platform for this ML/AI-focused freedom and responsibility! - you pointed out a few of our key competitors/cooperators like i have a list of about 20 more that i keep an eye on each and every day. i'm in close talks with all of them. - we're looking to partner with guys like Domino Data Labs who have a weak deployment story. - and we're constantly sharing experience and code with and and others. - we're super performance-focused, as well. we have a couple efforts going on including PMML optimization, native code generation, etc. - also super-focused on metrics and monitoring - including production-deployment dashboards targeted to data scientists. - i feel like our main competitors are actually the cloud providers. they're the ones that keep me awake. one of our underlying themes is to reverse engineer Google and AWS's Cloud ML APIs.  

Monday, January 16, 2017

CryptoNets: scoring deep learning on encrypted data

Last week I attended  an interesting lecture by Ran Gilad Bachrach from MSR. Ran presented CryptoNets who was reported in ICML 2016. CryptoNets allows to score trained deep learning models on encrypted data. They use homomorphic encryption a well known mechanism which allows computing encrypted products and sums. So the main trick is to limit the neural net operations to include only sums and products. To overcome this problem CryptoNet is using the square function as the only non-linear operation supported (vs. sigmoids, ReLU etc.)

On the up side, CryptoNets reports 99% accuracy on MNIST data which is the toy example everyone is using for deep learning. On the downside, you can not train a network but just score on new test data. Scoring is quite slow - around 5 minutes, although you can batch up to a few thousands scoring operations together at the same batch. Due to increasing complexity of the represented numbers the technique is also limited to a certain number of network layers.

I believe that in the coming few years additional research effort will be invested for trying to tackle the training of neural networks on private data without revealing the data contents.

Anyone who is interested in reading about other primitives who may be used for performing similar computation is welcome to take a look at my paper: D. Bickson, D. Dolev, G. Bezman and B. Pinkas Secure Multi-party Peer-to-Peer Numerical Computation. Proceedings of the 8th IEEE Peer-to-Peer Computing (P2P'08), Sept. 2008, Aachen, Germany - where we use both homomorphic encryption but also Shamir Secret Sharing to compute a similar distributed computation (in terms of sums and products).