Hot Takes | Cold Proofs

Every research culture has a version of the same problem. The people with the least power are the ones who most need to be in the room, and they are the first thing cut when the diary is full.

This episode is about what a university does about that. Dr Livia Lantini coordinates UWL's Early Career Researcher Network, which began as a conversation with the head of school: Professor Phil Cox asked why there wasn't something for people at the start of their careers, and she went and built it. It started in the School of Computing and Engineering, then spread. Six schools and colleges are in it now.

Professor Fabio Tosti directs the Faringdon Research Centre, and his answer to how it came to lean on early career researchers is that it never didn't. He was one himself when the centre was founded; Livia joined it as a PhD student. His statistic, offered reluctantly — he says he doesn't much like going by statistics — is that more than seventy per cent of the people contributing to Faringdon's research are formally early career researchers. Not assisting on it. Contributing to it, in leading positions, across the centre's disciplines.

The most useful advice in the episode is also the least glamorous. Livia's recommendation to any institution is to put people in a room and let them talk, because the person who could help you most — as a collaborator, as a mentor — is often sitting next to you or one door along, and neither of you knows what the other does. They have watched projects start exactly that way, between two people who met at an ECR meeting. "Sometimes being an academic is a solitary journey," Nino says, and nobody rushes to disagree.

Then there is the sharper version of the same argument. Representation is the thing people forget, Livia says: if you are not there and you hear it secondhand, you are not learning, you are not growing, and you are not representing your view. Fabio makes it structural. He has just been inaugurated as president of one of the EGU's divisions — a union whose annual assembly in Vienna draws more than twenty thousand scientists across five days — and the first item on his agenda there is the same as the one at UWL. If there is no representation of the young generation, there is no continuation. When a mandate ends, somebody has to already be ready to take it.

There is a definition worth stealing, too. The UK convention for "early career" is seven or eight years past the PhD. UWL made it ten, because people arrive from industry, or take time out for parental responsibilities, and a clock that starts at the viva punishes both. The other rule is a ceiling rather than a floor: senior lecturer at most. Not because associate professors stop needing support, but because the ECR groups are led by ECRs, and learning how to lead is the whole point. One group leader in another school, on being promoted, handed the role to someone more junior on purpose — and then taught her how to do it.

The ending is the reason to listen. Livia describes being a postdoc with no permanent job, a first teaching load, no idea how to manage her time or how to apply for a grant. She was lucky enough to be able to walk up to Fabio and ask. Not everybody has this luck. The network exists so that the answer to "I don't know how to do this" can be a group of people who say: we got you.

Also in this episode: the Faringdon Journal Club as collective supervision, where PhD students argue with senior academics at the same level; why getting a PhD student into a meeting with a council or an international partner is worth more than a report about it afterwards; diversity as a team-building problem rather than a compliance one; and whether a sector being told to do more with less can afford to treat early career researchers as the overhead rather than the answer.

Hosted by Atiyeh Ardakanian and Nino Auricchio, from the School of Computing and Engineering at the University of West London.

Professor Julie Wall and Dr Nasim Dadashi Serej on where bias actually comes from, and what you can do about it

Nobody in this episode claims you can build an AI with no bias in it. The interesting question is what you do once you accept that.

Dr Nasim Dadashi Serej's answer runs through the whole episode as one long metaphor: you train a model the way you raise a child. You feed it your own opinions and your own insights, you reward what you like and discourage what you don't, and then you are surprised when it comes back sounding like you. Reinforcement learning, she points out, is that analogy made literal — the agent explores, and you hand out rewards or punishments. Which prompts the obvious question from the hosts: so how do we punish ChatGPT?

Professor Julie Wall approaches it from the other end. A speech recognition system trained mostly on one accent will be better at that accent — that is not a scandal, it is what training data does. The work is in knowing the bias is there and deciding what you are willing to trade for it: a model tuned for equal error rates across categories will usually be less accurate overall than one tuned for accuracy alone, and that is a choice somebody has to make on purpose. She also corrects a premise mid-episode, politely and completely: these models were not born in Silicon Valley. AI goes back over eighty years, to the Second World War.

The healthcare section is the one that stays with you. In supervised learning somebody has to label the data, and in medical imaging that somebody is a clinician — so the model inherits not just their expertise but their off days. Nasim's term for it is intra-observer error: the same doctor, same case, different answer, because they are tired or in a low mood and want to get to the end of the list. Between two clinicians it becomes inter-observer error. Her argument for AI is not that it is smarter than a doctor; it is that it gives the same answer every time.

And a distinction worth stealing: the industry says "responsible AI" rather than "ethical AI" because ethics are not the same in every country a company operates in. GDPR works here. It does not travel.

Also in this episode: whether language models have been trained into agreeing with us, what happens to guardrails when people treat them as a puzzle to solve, the uncomfortable question of who does the content-moderation work that makes a model safe, and why explainable AI is the thing that gives both of them hope — asking a model, as Nasim puts it, the same thing you would ask a child: how did you get to that answer?

Chapters

• 00:00 Welcome, and why bias this time
• 00:27 What is bias?
• 01:39 Can you get rid of it?
• 02:30 Models trained around the world, with different ideologies
• 03:16 Accents, dialects and low-resource languages
• 04:38 The useful kind of bias
• 05:38 How do you detect it?
• 07:42 Reinforcement learning, and raising a child
• 09:40 Have these models been trained to agree with us?
• 11:41 Healthcare: whose labels are you learning from?
• 15:18 Accuracy or representativeness — the trade-off
• 16:26 Pain, gender and ethnicity in the clinic
• 17:58 Intra-observer and inter-observer error
• 20:13 "I don't think the models started in Silicon Valley"
• 22:41 Guardrails, and people who treat them as a game
• 25:11 Who does the moderation work
• 26:24 Responsible AI, not ethical AI
• 27:06 The same prompt in a different country
• 29:31 What gives them hope: explainable AI
• 31:28 Inside CAINT
• 33:46 Close

Hosted by Atiyeh Ardakanian and Nino Auricchio, from the School of Computing and Engineering at the University of West London.

Produced and edited by Jack Hoggard. Recorded at LSFMD, University of West London, with Tim Bowerman and Arthur Page.

#ArtificialIntelligence #AIBias #ResponsibleAI #ExplainableAI #MachineLearning #AIinHealthcare #AIEthics #NLP #Research #HigherEducation #Podcast #UWL #UniversityOfWestLondon

Professor Fabio Tosti, the director of UWL's Faringdon Research Centre on how you survey a bridge, a road or a ruin without laying a finger on it, why Heathrow behaves like a city in miniature, and the question his team is still trying to answer — whether the communities living with that infrastructure can actually use any of the data.You can learn a great deal about a bridge without drilling into it. Ground-penetrating radar sends electromagnetic waves into the subsurface and reads what comes back — moisture, cracks, cavities, decay — in much the same spirit as an X-ray, and without touching the structure at all.Professor Fabio Tosti directs The Faringdon Research Centre at the University of West London. It started as a ground-penetrating radar centre; it now runs interferometric radar, satellite remote sensing, embedded sensors, robotics and immersive technology, and Fabio would rather call it a centre for smart technology and society. His teams work with Ealing Council and the London boroughs, with Historic England on how ruined structures are damaged by environmental stressors, and with the Royal College of Art on where science, technology and art meet.In this episode: what non-destructive testing actually means and where it stops being remote sensing, why Heathrow is best understood as a city in miniature, how you explain a radar output to someone who has never seen one, and the research question with no precedent behind it — can high-spec data be turned into something a community can genuinely use?Hosted by Atiyeh Ardakanian and Nino Auricchio, from the School of Computing and Engineering at the University of West London.

Daniel Tweddle and Dr Eugenio Donati on accents, AI and who gets understood.

Speech technology has a data problem. Almost every benchmark dataset used to train systems that assess pronunciation is American English — so if you speak with a Geordie, Glaswegian or Brummie accent, the technology was never built with you in mind.

Daniel Tweddle is building the British English speech datasets that don't exist yet, annotating every mispronunciation by hand. His supervisor, Dr Eugenio Donati, has been through the same thing — he collected his PhD data during lockdown by recording himself.

In this episode: what mispronunciation detection and diagnosis actually means, why a voice-activated lift can't understand a Scottish accent, whether pronunciation training can target a dialect you choose rather than a "correct" one, and the question Daniel gets asked more than any other — who gets to decide what the right way to speak is?

Hosted by Atiyeh Ardakanian and Nino Auricchio, from the School of Computing and Engineering at the University of West London.

Chapters

• 00:00 Welcome
• 00:24 How Eugenio and Daniel got into sound
• 04:39 Favourite mispronunciation stories
• 06:36 What mispronunciation detection and diagnosis is
• 07:11 Britain has no speech data for this
• 09:20 Building the datasets from scratch
• 11:42 The word nobody could say
• 12:33 Regional accents and dialects
• 15:02 Pronunciation training for a dialect you choose
• 17:46 The lift that can't understand your accent
• 19:15 Detection vs diagnosis
• 21:20 Where the research goes next
• 23:05 Annotating vowels by hand
• 26:04 Why you sound like you
• 28:35 Accent, identity, and who decides
• 31:17 Deepfakes and voice as a health marker
• 33:22 Supervisor and PhD student
• 35:17 Life outside the research
• 38:02 Doing a PhD through lockdown

Produced and edited by Jack Hoggard. Recorded at LSFMD, University of West London, with Tim Bowerman and Arthur Page.

Hashtags: #SpeechTechnology #AI #Accents #Linguistics #Phonetics #MachineLearning #SpeechRecognition #PhDLife #HigherEducation #Podcast #UWL #UniversityOfWestLondon

The industry says "responsible AI". Prof. Julie Wall thinks that word was chosen carefully — ethics differ by market, and the rebrand quietly drops the part that binds you.

Professor Julie Wall and Dr Nasim Dadashiserej get into what that looks like in practice: deepfakes that have stopped being a celebrity problem and become something ordinary people use on each other, AI agents whose upside is real and whose blast radius is enormous, and regulation that is always, structurally, written for the technology we have already left behind.

Nasim's route into healthcare AI runs through six hours in A&E with her son, one nurse, and a single judgement call — an argument for decision support that no whitepaper makes as well.

(00:00) Introduction
(02:21) Six hours in A&E: why an engineer moved into healthcare AI
(04:38) "It should be called ethical AI"
(09:44) The more advanced the technology, the more vital the hand of the musician
(13:30) Regulation is always written for the last technology
(16:04) A teacher deepfaked his own head teacher
(20:41) Hundreds of agents, one blast radius

Hosted by Atiyeh Ardakanian and Nino Auricchio.
Hot Takes | Cold Proofs is produced by the School of Computing and Engineering at the University of West London.

A computing dean and a psychologist on why technology keeps getting built for a "typical user" who doesn't exist — and what happens when AI averages out the very people it was meant to serve.

Prof. Phil Cox (Dean of the School of Computing and Engineering) reads AI research proposals, and he has a blunt verdict on a lot of them: the coding is excellent, the understanding of people is absent. Prof. Jen Mayer (Head of the School of Human and Social Sciences), a psychologist, has the mirror-image problem — the outliers her field exists to study are exactly the ones a model is trained to discard.

Between them they get at something that goes well past AI: what we lose when we optimise for the average, and why the interesting cases are always the ones that get smoothed away.

Also in this episode: what happened when a dean built an AI avatar of himself from a one-minute recording and quietly ran it past his students, and an extraordinary food-engineering case study about what happens when brilliant technology meets actual human behaviour.

(00:00) Introduction
(04:24) "You don't understand anything about humans"
(11:59) The fat-free crisp that human behaviour defeated
(16:22) AI gives you the average — the outliers disappear
(23:45) A dean clones himself with AI

Hosted by Atiyeh Ardakanian and Nino Auricchio.
Hot Takes | Cold Proofs is produced by the School of Computing and Engineering at the University of West London.