My latest immersion into Beej's guide was through the GIOS course too! It also opened my eyes to the amount of familiarity/expertise my older co-workers had with systems programming (they helped me when when they heard me whining about the socket programming assignment). It taught me not to dismiss someone just because they don't know the latest TensorFlow API/Neurips paper.
From the introduction:
"""
Mostly it's a testimony that it's possible to have a successful career in computer science without taking military funding. My position has its roots in the Vietnam War, when I was a conscientious objector, did alternative service instead of submitting to the draft, and joined the Society of Friends (Quakers). During the 1980s and 90s, the position seemed to lose some of its urgency, so it became more of a testimony about career paths.
Since September 11, 2001, all the urgency is back. The defense of our country is at stake, so this testimony becomes critical. In short, I believe that non-violent methods of conflict resolution provide the only methods for protecting our country against the deadly threats we face in the long run. Military action, with its inevitable consequences to civilian populations, creates and fuels deadly threats, and therefore increases the danger that our country faces.
"""
(Copy-pasting the message I already posted in this thread. Seems more relevant here)
I think that adoption of privacy preserving data aggregation/analysis will become the norm. The most immediate applications are 1) telemetry data that is used for monitoring (for example, Google Chrome uses differential privacy for collecting this data), and, 2) services like Google Maps and Tinder-like dating apps. In these applications, essential user information can be represented as integer/boolean values (is user present in location X? True or False. how old is the user? device CPU usage right now? ...)
Based on my limited understanding* of differential privacy, it falls short on exactness (of aggregate values) and robustness (against malicious clients). I've lately been studying the literature on function secret sharing and I think it is a better alternative to DP. Take this paper: https://www.henrycg.com/files/academic/pres/nsdi17prio-slide....
Prio: Private, Robust and Scalable Computation of Aggregate Statistics
Data collection and aggregation is performed by multiple servers. Every user splits up her response into multiple shares and sends one share to each server. I've understood how private sums can be computed. Let me explain it with a straw-man scheme.
Example (slide 26):
x_1 (user 1 is on Bay Bridge):- true == 1 == 15 + (-12) + (-2)
x_2 (user 2 is on Bay Bridge):- false == 1 == (-10) + 7 + 3 ...
If all users send shares of their data to the servers in this manner AND as long as at least one server doesn't reveal the identities of the people who sent it responses, the servers can exchange the sum of the shares they've received. Adding the three responses will allow the servers to infer that there are _ number of users on Bay Bridge without revealing their identities.
This system can be made robust by using Secret-shared non-interactive proofs (SNIPs). This allows servers to test if Valid(X) holds without leaking X.
The authors also bring up the literature on computing interesting aggregates using private sums: average, variance, most popular (approx.), min and max (approx.), quality of regression model R^2, least-squares regression, stochastic gradient descent.
Bottom line: I found the discussion on deployment scenarios very interesting. Data servers with jurisdictional/geographical diversity, app store-app developer collaborations for eliminating risk in telemetry data analysis, enterprises contracting with external auditors for analyzing customer data, etc.
* - I understand the randomized response and, to some extent, the RAPPOR technique (used for collecting Chrome telemetry data) but the other literature in that community goes over my head.
* * - This technique is a black box to me at the moment.
I remember this paper on ad intelligence I read a few weeks ago: "Exploring ADINT: Using Ad Targeting for Surveillance on a Budget — or — How Alice Can Buy Ads to Track Bob".
I think that adoption of privacy preserving data aggregation/analysis will become the norm. The most immediate applications are 1) telemetry data that is used for monitoring (for example, Google Chrome uses differential privacy for collecting this data), and, 2) services like Google Maps and Tinder-like dating apps. In these applications, essential user information can be represented as integer/boolean values (is user present in location X? True or False. how old is the user? ...)
Based on my limited understanding* of differential privacy, it falls short on exactness (of aggregate values) and robustness (against malicious clients). I've lately been studying the literature on function secret sharing and I think it is a better alternative to DP.
Take this paper: https://www.henrycg.com/files/academic/pres/nsdi17prio-slide....
Prio: Private, Robust and Scalable Computation of Aggregate Statistics
Data collection and aggregation is performed by multiple servers. Every user splits up her response into multiple shares and sends one share to each server. I've understood how private sums can be computed. Let me explain it with a straw-man scheme.
Example (slide 26):
x_1 (user 1 is on Bay Bridge):- true == 1 == 15 + (-12) + (-2)
x_2 (user 2 is on Bay Bridge):- false == 1 == (-10) + 7 + 3 ...
If all users send shares of their data to the servers in this manner AND as long as at least one server doesn't reveal the identities of the people who sent it responses, the servers can exchange the sum of the shares they've received. Adding the three responses will allow the servers to infer that there are _ number of users on Bay Bridge without revealing their identities.
This system can be made robust by using Secret-shared non-interactive proofs (SNIPs). This allows servers to test if Valid(X) holds without leaking X.
The authors also bring up the literature on computing interesting aggregates using private sums: average, variance, most popular (approx.), min and max (approx.), quality of regression model R^2, least-squares regression, stochastic gradient descent.
Bottom line: I found the discussion on deployment scenarios very interesting. Data servers with jurisdictional/geographical diversity, app store-app developer collaborations for eliminating risk in telemetry data analysis, enterprises contracting with external auditors for analyzing customer data, etc.
* - I understand the randomized response and, to some extent, the RAPPOR technique (used for collecting Chrome telemetry data) but the other literature in that community goes over my head.
* * - This technique is a black box to me at the moment.
1. A log key anomaly detection LSTM model (a type of recurrent neural network) predicts the next log message type (log key). If it is wrong for any log entry, the operator is alerted that there is an anomaly in the execution flow.
2. If there is no anomaly in the execution flow, the parameter value and performance anomaly detection model is invoked. If the predicted parameter vector for that particular log key is very different from the actual parameter vector, the operator is alerted. The second model performs regression, in high-dimension space of the parameter feature vector. Therefore, the mean-squared error (MSE) is used as the loss function when training this model and when deciding to alert the operator.
What you are describing sounds similar to a "quine". A quine is a non-empty computer program which takes no input and produces a copy of its own source code as its only output.
MacroBase's pipeline is broken up into the following operators: Transform, Classify, Explain.
I find the Explanation operator very valuable and haven't seen something like this in any work in monitoring/anomaly detection (correct me if I'm wrong). It finds attribute value combinations that are disproportionately concentrated in the outlier datapoints. Outliers are determined by the classifier by looking at metric value(s). The classifier can be based on static thresholds, percentiles, unsupervised/supervised learning algorithms, etc.
Example: "A mobile application manufacturer issues a MacroBase query to monitor power drain readings (i.e., metrics) across devices and application versions (i.e., attributes). MacroBase's default operator pipeline reports that devices of type B264 running application version 2.26.3 are sixty times more likely to experience abnormally high power drain than the rest of the stream, indicating a potential problem with the interaction between devices of type B264 and application version 2.26.3".
I think this sort of engine/system is valuable for detecting and isolating anomalies in telemetry data.
Based on my limited understanding* of differential privacy, it falls short on exactness (of aggregate values) and robustness (against malicious clients). I've lately been studying the literature on function secret sharing and I think it is a better alternative to DP.
Prio: Private, Robust and Scalable Computation of Aggregate Statistics
Data collection and aggregation is performed by multiple servers. Every user splits up her response into multiple shares and sends one share to each server. I've understood how private sums can be computed. Let me explain it with a straw-man scheme.
Example (slide 26):
x_1 (user 1 is on Bay Bridge):- true == 1 == 15 + (-12) + (-2)
x_2 (user 2 is on Bay Bridge):- false == 1 == (-10) + 7 + 3
...
If all users send shares of their data to the servers in this manner AND as long as at least one server doesn't reveal the identities of the people who sent it responses, the servers can exchange the sum of the shares they've received. Adding the three responses will allow the servers to infer that there are _ number of users on Bay Bridge without revealing their identities.
This system can be made robust by using Secret-shared non-interactive proofs (SNIPs). This allows servers to test if Valid(X) holds without leaking X.
The authors also bring up the literature on computing interesting aggregates using private sums: average, variance, most popular (approx.), min and max (approx.), quality of regression model R^2, least-squares regression, stochastic gradient descent.
Bottom line: I found the discussion on deployment scenarios very interesting. Data servers with jurisdictional/geographical diversity, app store-app developer collaborations for eliminating risk in telemetry data analysis, enterprises contracting with external auditors for analyzing customer data, etc.
* - I understand the randomized response and, to some extent, the RAPPOR technique (used for collecting Chrome telemetry data) but the other literature in that community goes over my head.
* * - This technique is a black box to me at the moment.
Can anyone here offer advice for doing anomaly detection in distributed systems?
I'm not looking for advice on which models to use, per se. I'm more interested in how to go about things as a single-person team (building data warehousing infrastructure, etc.)
>> Other people do not expect this because there are papers about how to incentivize neurons to correspond to interesting features.
Could you clarify that statement? Are you saying that it was unusual for this group to find such a neuron? Also, I did not know that there are papers on how to incentivize neurons to correspond to interesting features. Could you please give me some references on those?
Apparently, DP has some detractors. I was told by my signal processing professor that differential privacy wasn't really a solution for privacy preserving data analysis. He said something along these lines: "if I know something about the underlying data distribution (Gaussian, etc.), it is possible to wash out the randomness."
Now, I don't understand DP well enough and information theory/signal processing still seems a bit like "dragons be here" to me. But, I want to take a stab at trying to reason why he said that.
For example, take randomized response (the only DP technique I understand). That is vulnerable to a longitudinal attack: a person can query repeatedly to wash out the randomness. If you think about it, isn't it the almost the inverse of a repetition code (error correction)? There, you're trying to use redundancy (repetition) to remove noise.
>>>Sentimental neuron sounds fascinating too. I didn't realize individual neurons could be talked about or understood outside of the concept of the NN. I am thinking in terms of "black box" its often referenced to in some articles.
Yes, I agree. I recall seeing such individual neuron analysis before in Karpathy's "The Unreasonable Effectiveness of Recurrent Neural Networks". He takes a char-rnn that was training to predict the next character for source code and finds neurons that have learned to do paranthesis/bracket opening/closing.
(Apologies for the slightly incoherent post below)
I've been noticing a lot of work that digs into ML model internals (as they've done here to find the sentiment neuron) to understand why they work or use them to do something. Let me recall interesting instances of this:
1. Sander Dieleman's blog post about using CNNs at Spotify to do content-based recommendations for music. He didn't write about the system performance but collected playlists that maximally activated each of the CNN filters (early layer filters picked up on primitive audio features, later ones picked up on more abstract features). The filters were essentially learning the musical elements specific to various subgenres.
2. The ELI5 - Explain Like I'm Five - Python Library. It explains the outputs of many linear classifiers. I've used it to explain why a text classifier was given a certain prediction: it highlights features to show how much or little they contribute to the prediction (dark red for negative contribution, dark green for positive contribution).
3. FairML: Auditing black-box models. Inspecting the model to find which features are important. With privacy and security concerns too!
Since deep learning/machine learning is very empirical at this stage, I think improvements in instrumentation can lead to ML/DL being adopted for more kinds of problems. For example: chemical/biological data. I'd be highly curious to what new ways of inspecting such kinds of data would be insightful (we can play audio input that maximally active filters for a music-related network, we can visualize what filters are learning in an object detection network, etc.)
"Each time you make a Voice Call on Telegram, a neural network learns from your and your device‘s feedback (naturally, it doesn’t have access to the contents of the conversation, it has only technical information such as network speed, ping times, packet loss percentage, etc.). The machine optimizes dozens of parameters based on this input, improving the quality of future calls on the given device and network."