I learned about shroedingers cat from watching The Big Bang Theory. I also am covered in tattoos...lol your mom would probably think I am a heathen. My tattoos are not nerdy. I have a drawing that my son did as an assignment in his 3rd grade class that included some zombies so his teacher did not want to put it on the wall with the other kids' drawings because she thought it might scare them....so I put it on my arm. I also have a drawing my grandmother drew for me when I was in middle school along with a note that she wrote saying "I love you" in her writing. It might be time to do a nerdy tatt, but it may be something like my character from the video game world of warcraft :)
Hi there , it would be really cool if you covered the paper ,” Robust feature level adversaries are interpretability tools “ klab.tch.harvard.edu/publications/PDFs/gk8093.pdf
Looking forward to the video when you explain how to implement the 3 ways you mentioned on how to defend from those attacks. Thanks for the video as usual!
I think of machine learning as the evolving extrapolation. To clearfy, remember search back in the days. We had to write everything right down to the letter. Otherwise, it won't show up in the search results. Now computers have improved in extrapolation and can understand what we're searching for even if we misspelled. And so the evolving of extrapolation continues with ML and AI.
I think that most of the mistakes that's happening with AI originates from humans trying to control it. If a multimodel is in control of itself, it will behave logically. But people want to be in control and that's the source of instability. They basically kids argue about who gets to drive the car.
@Keiranful I looked it up, and from what understand, transformers do un-sequential processing while convolutional do sequential. Both ML algorithms can do the same things, but due to the inherent technicalities, they excel in different things.
It's always problematic to have a model released for inference without a contiguous fine tuning plan. Especially open source models. It makes all the models deployed identical, and since the model weights are available online for anyone, it's a matter of time until a vulnerability is found and used. So constant fine tuning is absolutely necessary even if it doesn't improve the performance of the model significantly.
Great episode once again! Couple of questions. The noise an adversary would make if they were to change the image or text to something false compared to a predetermined result the adversary chooses, would it be a big difference? And are there ways to tell on the owners end that manipulation has taken place? Also, when building out a new ai design, apart from implementation of correct security controls, are there any ways to track issues like hallucinations/data poisoning, or would that not be possible until the audit phase?
Great questions! The difference between the noise patterns are usually quite subtle (even imperceptible) to humans, but still make a big difference to the model. And yes, there are ways the system can spot an adversarial attack (things like input testing/validation) or prevent them from working (training the models with adversarial examples). The challenge is actually making these kinds of tests the norm, rather than an afterthought (or not done at all, which is still too often the case). I love the second part of your question as well because it speaks to the maturity of the data science process - yes there are definitely ways models can be built to be 'secure from the start', much like software systems are built these days, however it would require a fundamental change in the model-building a verification process - not a technical challenge but a process/human challenge - and potentially even more challenging!
Haha well spotted! I do actually (just a minor opportunity) but it's something I'd love to do more of.. if only to see AI security taken more seriously by seeing it in mainstream media. Also who I am to turn down a free makeup/hair session 🙃
Hi Harriet, Actually, Einstein first came up with the photoelectric effect which is what he won his Nobel prize for. I remember doing both SR and GR at Monash (not in physics but in maths) and what is far more powerful is to write the energy-mass equivalence equation using speed of light units so that it actually reads: E = m. This struck me very powerfully at the time because it actually states right out that energy and mass are exactly the same. I recognised Schroedinger's equation of course but there are always Maxwell's equations which are really great too. I love how Maxwell's equations look almost symmetric but not quite always leaving a gap so that there is something which we don't know yet. I also did magnetohydrodynamics and there are a nice bunch of fluid and EM equations in that, really amazing physics in that but not easy to understand. I loved your background, how you worked at different places, all of this is really good for your CV. You said you had also done a Masters degree? Was that on the bioanthropology or is that undergraduate. You would have loved doing an honours degree, all that extra graduate level knowledge. I did mine at Monash long ago now. Maybe you could look for some research type work all over the world although they would most likely prefer you having a PhD. But, maybe you could work in engineering style research which is what I did for a while. It's great to see someone else doing interesting things. All the best, Frank
Thanks for taking the time to write this reply Frank! It's always great to hear from other people working on interesting things. I have a PhD in machine learning and run my own AI business and don't plan on working for anyone else again.. and that's the benefit of being able to connect with likeminded people on this platform!
@@HarrietHacks Dear Harriet, thank you for your reply. I can see that you have chosen the right path with the business and perfectly understand that working for yourself is far superior to most other forms of work. I wish you great success in your endeavours, I have no doubt you are already doing and will continue to do well. I once thought along similar lines as you do now although given my own background, I did not have expertise in such an area, my own PhD being in using mathematical models in solving complex multiphase flow problems in engineering research. But then life is full of choices isn't it. Good luck and enjoy, All the best, Frank.
I think a lot of people get confused in the difference between sentience, sapience, conscious, self-aware, and (super/artificial general) intelligence. You can have something that could do any task a human could do, but is incapable of acting or "thinking" on its own. And a lot of people miss that sort of nuance.
This is an important point that many miss, AGI is not the same as you or I, it is just a system that is as capable as an average human at a wide range of tasks. sentience is not a requirement and I think we are a lot closer to AGI and still a fair way off sentience. The fact that something can pass the turing test just means that now that the turing test has itself been tested it is time for a better test
@@HarrietHacks I agree. I personally think if you gave the current SOTA vision enabled chatbots to AI researchers 50 years ago, easily 90% of them would agree that it is AGI, Even if you explained the current limitations.
Another great vid! Regarding the learning curve for future ai professionals, do you think free content is valuable/informative enough for enabling possible job applications or are paid courses at a point that they are necessary to attempt the jump into the ai space. Also, extending on that, do many ai focused positions exist yet or is it like when cyber security became a thing and the person who put their hand up to install antivirus from the IT team became the resident cyber security person.
That's a great couple of questions - I personally think it depends what future AI related role a person wants to have (and also what is required by the market). I remember back in the day, data scientists were in such demand that a short online course was enough to get someone hired.. now it's much more competitive. And those people probably could have benefitted from more in-depth training to future-proof both their jobs and the models they are creating! And to your other question, I'm seeing both ends of that spectrum - some organisations (usually larger, more technical) have AI roles, and others are largely volunteer-based. I'm interested to know what everyone else here thinks too!
Fantastic Episode! Looking forward to more in this series, really informative, and the effects & delivery are great! And I think I'd say I'm cautiously optimistic about AI, but looking forward to expanding my understanding of topic, after all as Marie Curie once said, "Nothing in life is to be feared, it is only to be understood. Now is the time to understand more, so that we may fear less."
I have been eagerly waiting for this series and I thought there is gonna be a AI Joke lol, but I have one for you ..Why did the computer keep sneezing? It had a virus!
Haha that's a good one! Don't worry, there may not be so many jokes in the remastered series but there's no way the comedian in me can keep them out of my channel altogether! :)
I have been there last year, great place to chill and relax ,their traditon is amazing to watch,hope you had a good time and have you seen the monuments in Jakarta ?
I have already finished watching Intro to AI series ,the content is awesome and it have inspired me to explore more on this field ,im looking forward to the new series!!
I'm looking forward to seeing the the new series. Echoing what @Hydroidtv said audio is over saturated. If you have some kind of audio input interface please check the mic input. It's hard to listen at this level.
And I thought computer science was hard, this AV stuff is killing me! I feel like we're all in this together, thanks for the advice - it'll be fixed for the next episode :)
Awesome work on stepping up the production value. If you want to take your audio to the next level, turn down the gain on the microphone a little to eliminate the clipping. It will come out sounding a lot better!
@@HarrietHacks Just finished the 10 episodes and I really wanted to thank you as you opened my eyes to new possibilities! I wanted to work as DS/ML in the cybersecurity field but was not sure how to link it, and now... I know. Thanks a lot! Looking forward to 2024 new serie! Cheers from Mexico.
Thank you, I'm so glad to hear that! The new series will start coming out next week, there's a bit of content overlap to start off with but it's much better quality! :)
Looks like an amazing experience! Thanks for sharing! One of the reasons I changed careers to cyber security was for the freedom of remote work. This inspires me to work towards a job where I can see the world!
This is exactly what I needed with all of my pent-up frustration. Somehow, you always manage to post about exactly what I'm looking for. I've never been so in love with a channel. 🥰
Best joke yet. Thanks for breaking this down and making it accessible. I'm interested in the field of AI security but don't have the math background (yet). Coming from a cyber security background though and am excited to see where this field goes. Hopefully we can learn from some of our past mistakes and jump up the maturity ladder faster.
Thank you so much for this, Harriet! I have been interested in both cybersecurity and AI for a while. I went back and forth between studying in both fields. I recently learned there was a way to combine them, and I was very excited by the idea. However, I initially thought that since this field was still relatively new, the resources would be very limited. While they're definitely not as abundant as other fields in the tech industry, the sources that I have found have been so useful.
@@HarrietHacks I've discovered that when I go through the content, there are some terms, phrases, and concepts that I find difficult to understand. Could you please give me some advice on what fundamental topics would be good for me to learn?
@@HarrietHacks Specifically the math and programming skills I need to know. When you dived into the math behind the models in one of the videos of the series, I got confused. And I have been learning Python and building some new projects, but I don't see the correlation between how I would apply these skills. So, what specific concepts of math and programming should I focus on in order to build more in-depth knowledge?
I'd say linear algebra and calculus really underpin the mathematics behind machine learning, and most 'Intro to ML' courses should cover how to apply these. I know the channel 3Blue1Brown also does a good job of explaining the math x ML intersection (and definitely others I'm not aware of!). The app Brilliant also has some great visualisations for math and code that I find really helpful (although you have to pay for a subscription). But I would also say that it took me a long time to really understand what was going on, I definitely didn't understand it the first time I was exposed to it, and that's pretty normal (and ok!). I'm sure there must be some good 'math in plain english' videos out there.. otherwise maybe I should do some! Good luck, you'll get there!
Hi Harriet. Just wanted to say that I found your series on medium and really enjoyed it. I read them all at once and eagerly awaited the last few parts. Please do continue doing these! Cheers
Amazing series! Really appreciate all the recommendations for upskilling! Really hope you continue your RU-vid/blog journey, you're great at it! Looking forward to the lab!
This comment made my day! Thank you for saying that, it is really hard putting yourself out there on the internet but I want to keep doing it and getting better at it! :)
I used to do data security for a company and even led a pci-dss audit, it was a pain, but a good thing to do. I like this perspective on AI and security. Thanks for making these! I do feel like "misinformation" is still a bit too broad/unspecific a thing, like "fairness", because of the existence of exaggeration, developing stories, different perspectives, etc.
I like this. I'm from Thailand, In my country right now people on majority were so afraid of the "AI" thing, rather than just finding the origin of it and how to use it, or comprehend it. Love your content.